(The International Series in Engineering and Computer Science 874) Kathleen Philips, Arthur H. M. van Roermund (auth.) - ΣΔ A - D CONVERSION FOR SIGNAL CONDITIONING-Springer Netherlands (2006)

Σ∆ A/D CONVERSION FOR SIGNAL CONDITIONING
THE INTERNATIONAL SERIES IN ENGINEERING AND COMPUTER

tekst
SCIENCE
ANALOG CIRCUITS AND SIGNAL PROCESSING

Consulting Editor: Mohammed Ismail. Ohio State University
Related Titles:
WIDE-BAND WIDTH HIGH-DYNAMIC RANGE D/A CONVERTERS

Doris, Konstantinos, ven Roermund, Arthur, Leenaerts, Domine
Vol. 871, ISBN: 0-387-30415-0
METHODOLOGY FOR THE DIGITAL CALIBRATION OF ANALOG CIRCUITS AND
SYSTEMS: WITH CASE STUDIES
Pastre, Marc, Kayal, Maher
Vol. 870, ISBN: 1-4020-4252-3
HIGH-SPEED PHOTODIODES IN STANDARD CMOS TECHNOLOGY
Radovanovic, Sasa, Annema, Anne-Johan, Nauta, Bram
Vol. 869, ISBN: 0-387-28591-1
LOW-POWER LOW-VOLTAGE SIGMA-DELTA MODULATORS IN NANOMETER CMOS
Yao, Libin, Steyaert, Michiel, Sansen, Willy
Vol. 868, ISBN: 1-4020-4139-X
DESIGN OF VERY HIGH-FREQUENCY MULTIRATE SWITCHED-CAPACITOR
CIRCUITS
U, Seng Pan, Martins, Rui Paulo, Epifânio da Franca, José
Vol. 867, ISBN: 0-387-26121-4
DYNAMIC CHARACTERISATION OF ANALOGUE-TO-DIGITAL CONVERTERS
Dallet, Dominique; Machado da Silva, José (Eds.)
Vol. 860, ISBN: 0-387-25902-3
ANALOG DESIGN ESSENTIALS
Sansen, Willy
Vol. 859, ISBN: 0-387-25746-2
DESIGN OF WIRELESS AUTONOMOUS DATALOGGER IC'S
Claes and Sansen
Vol. 854, ISBN: 1-4020-3208-0
MATCHING PROPERTIES OF DEEP SUB-MICRON MOS TRANSISTORS
Croon, Sansen, Maes
Vol. 851, ISBN: 0-387-24314-3
LNA-ESD CO-DESIGN FOR FULLY INTEGRATED CMOS WIRELESS RECEIVERS
Leroux and Steyaert
Vol. 843, ISBN: 1-4020-3190-4
SYSTEMATIC MODELING AND ANALYSIS OF TELECOM FRONTENDS AND THEIR
BUILDING BLOCKS
Vanassche, Gielen, Sansen
Vol. 842, ISBN: 1-4020-3173-4
LOW-POWER DEEP SUB-MICRON CMOS LOGIC SUB-THRESHOLD CURRENT
REDUCTION
van der Meer, van Staveren, van Roermund
Vol. 841, ISBN: 1-4020-2848-2
WIDEBAND LOW NOISE AMPLIFIERS EXPLOITING THERMAL NOISE
CANCELLATION
Bruccoleri, Klumperink, Nauta
Vol. 840, ISBN: 1-4020-3187-4
CMOS PLL SYNTHESIZERS: ANALYSIS AND DESIGN
Shu, Keliu, Sánchez-Sinencio, Edgar
Vol. 783, ISBN: 0-387-23668-6
SYSTEMATIC DESIGN OF SIGMA-DELTA ANALOG-TO-DIGITAL CONVERTERS
Bajdechi and Huijsing
Vol. 768, ISBN: 1-4020-7945-1
OPERATIONAL AMPLIFIER SPEED AND ACCURACY IMPROVEMENT
Ivanov and Filanovsky
Vol. 763, ISBN: 1-4020-7772-6
STATIC AND DYNAMIC PERFORMANCE LIMITATIONS FOR HIGH SPEED
D/A CONVERTERS
van den Bosch, Steyaert and Sansen
Vol. 761, ISBN: 1-4020-7761-0
DESIGN AND ANALYSIS OF HIGH EFFICIENCY LINE DRIVERS FOR Xdsl
Piessens and Steyaert
Vol. 759, ISBN: 1-4020-7727-0
LOW POWER ANALOG CMOS FOR CARDIAC PACEMAKERS
Silveira and Flandre
Vol. 758, ISBN: 1-4020-7719-X
MIXED-SIGNAL LAYOUT GENERATION CONCEPTS
Lin, van Roermund, Leenaerts
VOLUME 595
Vol. 751, ISBN: 1-4020-7598-7
Σ∆ A/D CONVERSION FOR
SIGNAL CONDITIONING
by
Kathleen Philips
Philips Research Laboratories,
Eindhoven, The Netherlands
and
Arthur H. M. van Roermund

Eindhoven University of Technology,
The Netherlands
A C.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN-10 1-4020-4679-0 (HB)

ISBN-13 978-1-4020-4679-7 (HB)
ISBN-10 1-4020-4680-4 (e-book)
ISBN-13 978-1-4020-4680-3 (e-book)
Published by Springer,
P.O. Box 17, 3300 AA Dordrecht, The Netherlands.
www.springer.com
Printed on acid-free paper
All Rights Reserved

© 2006 Springer
No part of this work may be reproduced, stored in a retrieval system, or transmitted
in any form or by any means, electronic, mechanical, photocopying, microfilming, recording
or otherwise, without written permission from the Publisher, with the exception
of any material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work.
Printed in the Netherlands.

Contents
List of symbols and abbreviations ix
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 The signal conditioning channel 9

2.1 Generic communication channel . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Performance parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Conventional conditioning channels . . . . . . . . . . . . . . . . . . . . 10
2.4 Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.1 Technology advances . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4.2 System demands . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.3 Advances in digital signal processing and analog circuit design . . 14
2.4.4 Digitization of the architecture . . . . . . . . . . . . . . . . . . . 14
2.5 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 A/D conversion 21
3.1 Historical overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 State-of-the-art in A/D conversion . . . . . . . . . . . . . . . . . . . 22
3.2.1 Architectural considerations . . . . . . . . . . . . . . . . . . . . 23
3.2.2 Implementation aspects . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.3 Performance metrics for ADCs . . . . . . . . . . . . . . . . 28
3.3 ADCs in future conditioning channels . . . . . . . . . . . . . . . . . 29
3.3.1 The Shannon theorem and based signal conditioning . . . . . 29
3.3.2 Comparison of Nyquist and based signal conditioning . . . . 30
3.3.3 Survey of published power/performance values . . . . . . . . . . 32
3.4 Limitations of A/D conversion . . . . . . . . . . . . . . . . . . . . . 32
3.4.1 Linear limitations . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.2 Non-linear limitations . . . . . . . . . . . . . . . . . . . . . . . 33
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
vi CONTENTS
4 Power consumption in channel building blocks 37

4.1 Literature on power/performance analysis . . . . . . . . . . . . . . . . . 37
4.2 Figures-of-merit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.1 FOM related to thermal noise . . . . . . . . . . . . . . . . . . . 38
4.2.2 FOM including distortion . . . . . . . . . . . . . . . . . . . . . 39
4.2.3 FOM related to signal resolution . . . . . . . . . . . . . . . . . . 39
4.3 Power consumption in analog conditioning circuits . . . . . . . . . . . . 41
4.3.1 Power/performance relations . . . . . . . . . . . . . . . . . . . . 41
4.3.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.4 Power consumption in a ADC . . . . . . . . . . . . . . . . . . . . . 44
4.4.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.5 Power consumption in digital conditioning circuits . . . . . . . . . . . . 47
4.5.1 Filter functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.6 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5 Full-analog and full-digital conditioning channels 57

5.1 Full-analog conditioning channel . . . . . . . . . . . . . . . . . . . . . . 57
5.1.1 The conditioning channel . . . . . . . . . . . . . . . . . . . . . . 58
5.2 Full-digital conditioning channel . . . . . . . . . . . . . . . . . . . . . . 59
5.2.2 Power/performance analysis . . . . . . . . . . . . . . . . . . . . 61
5.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6 Conditioning ADCs 67
6.1 Generic conditioning ADC . . . . . . . . . . . . . . . . . . . . . . . 67
6.1.1 Concept of operation . . . . . . . . . . . . . . . . . . . . . . . . 67
6.1.2 Universal model of a modulator . . . . . . . . . . . . . . . . 72
6.1.3 Interferer immunity . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.2 Signal conditioning in the decimation filter . . . . . . . . . . . . . . . . . 79
6.3 Signal conditioning with a restricted filtering STF . . . . . . . . . . . . . 84
6.3.4 Conditioning hybrid ADC . . . . . . . . . . . . . . . . . . . 92
6.4 Signal conditioning by unrestricted STF design . . . . . . . . . . . . . . 94
CONTENTS vii

6.5 Comparison of conditioning ADCs . . . . . . . . . . . . . . . . . . . . . 99
6.5.1 Comparison of topologies . . . . . . . . . . . . . . . . . . . . . 100
6.5.2 Flexibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.5.3 Power consumption . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.5.4 Guidelines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7 Digitization of the inter-die interface 105

7.1 Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.2 Power in the interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.2.1 Analog interface . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.2.2 Digital interface after decimation . . . . . . . . . . . . . . . . . 108
7.2.3 Digital interface before decimation . . . . . . . . . . . . . . . . 108
7.2.4 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.3 Application to the conditioning channels . . . . . . . . . . . . . . . . . . 110
7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8 Highly analog and highly digital channels for FM/AM radio 113
8.1 System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.1.1 Conventional radio with analog demodulation . . . . . . . . . . . 114
8.1.2 Radio with digital demodulation . . . . . . . . . . . . . . . . . . 115
8.2 VGA design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8.2.1 Highly linear VGA design . . . . . . . . . . . . . . . . . . . . . 117
8.2.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8.3 ADC design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.3.1 Conventional solutions . . . . . . . . . . . . . . . . . . . . . . . 124
8.3.2 ADC with integrated passive mixer . . . . . . . . . . . . . . 125
8.3.3 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.4 Evaluation of the channel . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.4.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.4.2 Benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
9 Conditioning ADCs for Bluetooth 139

9.1 System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
9.1.1 Conventional radio with analog demodulation . . . . . . . . . . . 140
9.1.2 Radio with digital demodulation and analog signal-conditioning . 141
9.1.3 Radio with digital demodulation, without analog signal-conditioning141
9.2 Feed forward ADC . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
9.2.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
9.2.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.3 Conditioning feedback ADC . . . . . . . . . . . . . . . . . . . . . . 156
9.3.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
9.3.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
viii CONTENTS
9.4 FFB-ADC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

9.4.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.4.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
9.5 Evaluation of the channels . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.5.1 Benchmark with published ADCs . . . . . . . . . . . . . . . . . 182
9.5.2 Comparison of the presented ADCs . . . . . . . . . . . . . . . . 182
9.5.3 Benchmark with published Bluetooth conditioning channels . . . 187
9.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
10 General conclusions 191
A Overview of published ADCs 193
B Power/performance relation of analog circuits 199

B.1 Simple differential pair . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
B.2 Differential pair in a global feed-back configuration . . . . . . . . . . . . 200
B.3 Degenerated differential pair . . . . . . . . . . . . . . . . . . . . . . . . 201
C Power/performance relation of digital filters 203

C.1 Analysis of the filter topology . . . . . . . . . . . . . . . . . . . . . . . 203
C.2 Calculation of filter parameters . . . . . . . . . . . . . . . . . . . . . . . 204
C.3 Calculation of power consumption . . . . . . . . . . . . . . . . . . . . . 205
C.4 Extension to other implementations . . . . . . . . . . . . . . . . . . . . 206
D Third-order distortion in analog circuits and ADCs 207
E Power consumption in a data interface 211

E.1 Analog data interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
E.2 Digital data interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
References 215
List of symbols and abbreviations
Symbol Description Unit
BW bandwidth of the wanted signal Hz

C capacitance F
c linearized quantizer gain for a ADC
C0 (t) matched filter
CCh conditioning channel
d linearized DAC gain for a ADC
dBFS ratio (in decibels) of the signal power compared to the full- dBFS
scale power of a digital data format
DR dynamic range; i.e. ratio of the maximum signal power to the
minimum detectable signal power in the same bandwidth
DRdB DR expressed in decibels dB
DSP Digital Signal Processing
F implementation factor, i.e. the ratio of FOM(ADC) over
FOM(reference ADC)
fa transition frequency from a high-order to a first-order slope Hz
of a transfer function
FFB-ADC filtering-feedback ADC
fug unity-gain bandwidth Hz
fs Nyquist sample rate
fsw average switching frequency; i.e. the inverse of the average Hz
number of 0 −1 transitions in a digital data sequence
IM3 ratio of the amplitude of the third-order intermodulation
component, to the amplitude of the fundamental signals. The
fundamental signals both are applied at half of the full-scale
level.
IM3,dB IM3 expressed in decibels dBc
k Boltzmann’s constant equaling 1.38×10−23 J/K J/K
L order of the loop filter of a ADC
m over-sampling factor
MASH multi-stage noise shaping
N number of bits of a quantizer
x LIST OF SYMBOLS
Symbol Description Unit
NTF noise transfer function

p ratio between the bandwidth of the entire signal (including
the wanted and interferer channels) and the bandwidth of the
wanted signal only
P power consumption W
q ratio of the amplitude of an overall signal, consisting of both
wanted and interferer components, to that of the wanted sig-
nal only
SINAD signal/noise-and-distortion ratio; i.e. ratio of the power of the
wanted signal to the sum of the noise power and the distortion
power. All are integrated over the channel bandwidth and are
present simultaneously
SINADdB SINAD expressed in decibels dB
SNR signal/noise ratio; i.e. ratio of the power of the wanted sig-
nal to the noise power. Both are integrated over the channel
bandwidth and are present simultaneously
SNRdB SNR expressed in decibels dB
STF signal transfer function
T absolute temperature K
v̂IN instantaneous amplitude of a wanted signal V
v̂MIN minimum amplitude of a wanted signal V
v̂MAX maximum amplitude of a wanted signal V
vn short-hand notation for the rms value of an input-referred, V
equivalent noise source, measured over the wanted channel
only
VDD positive supply voltage V
VSS negative supply voltage V
Chapter 1
Introduction
1.1 Background
Moore’s Law predicts a decrease by a factor of two in the feature size of CMOS tech-
nology every three years and has been valid for years. It implies a doubling of the op-
eration speed and a four times higher transistor count per unit of area, every three years.
The combination leads to an eight times higher processing capability per unit of area.
This on-going miniaturization allows the integration of complex electronic systems with
millions of transistors (Very-Large-Scale-Integration) and enables the integration of elec-
tronic systems.
An electronic system
A generic picture of an integrated electronic system is shown in fig. 1.1. The heart of
the system is the signal processing core. This core supports a wide variety of functions,
such as customization and programmability of multiple applications, channel coding, the
definition of the user interface, etc. These functions are enabled by DSP, a controller CPU
and various blocks of memory. In advanced ICs these blocks provide (almost) all signal
processing and usually dominate in the overall power and area consumption of integrated
systems. The huge data rates involved, require high-speed busses for communication
between these blocks. A power-management unit fuels the system by providing the ap-
propriate supply voltages and currents.
Communication with the physical world is realized by a chain of mixed-signal, ana-
log and RF-circuitry. This chain acts as a “signal-conditioning channel”. It translates
physical signals into binary representations or vice versa. This channel also comprises
amplification, filtering and possibly frequency translation. It is the challenge of present
day mixed-signal and RF design to integrate the signal-conditioning channel at a low
power consumption and a high performance level.
1
2 INTRODUCTION
Figure 1.1: Electronic system including “signal-processing” units and some

example “signal-conditioning” channels for communication to the
physical world [1]
Digital signal processing
More and more digital systems and standards have been conceived: DECT has replaced
analog cordless systems, CD has overtaken the market of analog audio and digital cam-
eras are conquering conventional photography. These new standards offer higher quality
and more features thanks to the digital signal processing. The digital solution is moreover
easily scalable to new technologies and changing systems.
Clearly, in the signal-processing arena “full-digital” is becoming a fact for the ma-
jority of systems. A similar evolution is happening to the signal-conditioning channels
shown in fig. 1.1 although the feasibility and economics of “highly-digitized signal con-
ditioning” have not yet been proven.
Digitization of signal conditioning
Fig. 1.2 depicts a conventional (i.e. dominantly analog) implementation of an A/D and a
D/A type of signal-conditioning channel. In fact, this could be the block diagram detailing
any of the signal-conditioning channels introduced in fig. 1.1. These channels select the
wanted signal in the time and the frequency domain, amplify the wanted signal, suppress
interferer signals and noise, perform A/D or D/A conversion and de- or encoding of the
signal.
1.1. BACKGROUND 3
Figure 1.2: Block schematic of conventional A/D and D/A conditioning channel
The signal conditioning happens in either the analog or the digital domain1 . Pre-
dominant analog signal conditioning relaxes the bandwidth and the dynamic range re-
quirements of all the following blocks. Predominant digital conditioning reduces the
number of analog blocks needed, while the implementation of the digital blocks benefits
from Moore’s Law. The requirements imposed on the data converters, however, become
substantially stricter. As the ADC and DAC move towards the antenna a much higher
sample rate and significantly higher resolution and linearity are required (see example on
page 15).
While digital processing seems to come for free in advanced CMOS technologies,
any performance increase of analog circuitry leads to higher power consumption. More-
over, migration of an analog circuit to a new technology generation may imply a power
increase, even at constant performance requirements [3], [4]. Although some transistor
parameters have improved in advanced technologies, the negative effect of the decreasing
supply voltage has over-compensated this for generations beyond 0.25µm-CMOS [5].
Therefore, new circuit techniques need to be developed in order to bridge this perfor-
mance gap. Evolution in analog circuit techniques is, however, very slow in comparison
with the revolution that has taken place in the complexity of digital processing and algo-
rithms.
Obviously, digitization of the signal-conditioning channel in an advanced CMOS
technology, imposes a substantial burden on the analog circuits and on the data convert-
ers especially. All-digital signal conditioning is therefore not necessarily the best option
in view of the overall optimization of the signal-conditioning channel.
Digitization of inter-die interfaces

The signal-conditioning channel can be integrated on a single die or often it extends over
multiple dies (in order to take advantage of a dedicated technology or for reasons of stan-
dardization of the interface). In the latter case, digitization of the channel may lead to a
1 In this book, it is assumed that de- and encoding of the signal occurs in the digital domain. This is obvi-
ous for digital communication schemes but is also becoming the de-facto implementation for analog systems.
A good example are FM radio receivers in which analog demodulation -although adequate and low-cost- is
substituted by digital demodulation [2] featuring all the previously mentioned advantages.
4 INTRODUCTION
Figure 1.3: Nomenclature with respect to the signal-conditioning channel
new inter-die interface: as the ADC or DAC shifts towards the antenna a previously analog
interface may be replaced by a digital interface. Hence, the cost of an inter-die interface
will strongly depend on the degree of digitization of the signal-conditioning channel, and
must be included in the overall optimization of the cost/performance ratio of the channel.
1.2 Scope
This book studies the digitization of the signal-conditioning channel. In particular, it
focuses on the consequences of digitization on the power consumption of the channel
in relation to the performance target. The target “performance” is evaluated in terms of
noise, distortion and bandwidth. In generic terms, the aim of this book is to improve the
power/performance relation of the conditioning channel by balancing analog and digital
signal conditioning. This is pursued while striving for a highly-digitized solution. Some
limitations on the scope of the text are briefly motivated next.
Baseband A/D conditioning channels

The text focuses on signal conditioning in an A/D type of channel, operating at baseband.
The key circuits in the analysis are filters, variable-gain amplifiers and data converters. At
present, digitization of this (part of the) signal-conditioning channel -though very chal-
lenging still- is becoming feasible for various systems. On the contrary, power-efficient
digitization at the IF or RF frequency can be considered as a next -but further-off- step. In
addition, the de- and encoding of the signal are left out because the digitization of these
blocks has become a reality already.
The nomenclature as defined in fig. 1.3 will be used. “Conditioning channel” is used
as a shorthand notation for “signal-conditioning channel”. As explained above, it only
refers to the conditioning actions identified in fig. 1.3. Moreover, the term “signal con-
ditioning” only refers to filtering and variable gain or word-length scaling, without the
data conversion. In chapter 6, this functionality is integrated into a ADC. Then, the
terminology of “a conditioning ADC” is used.
Continuous-time single-bit sigma-delta conversion

The choice for sigma-delta data conversion is motivated by the evolution of CMOS tech-
nology and the system need for low power data conversion (see chapters 2 and 3) and by
1.2. SCOPE 5
the potential of sigma-delta converters for digitizing the conditioning channel in a power
efficient way. The latter argument is demonstrated throughout the book.
This choice does limit the analysis to channels with a “narrow” bandwidth. At
present, ADCs with a signal bandwidth up to 40MHz have been reported [6] in
CMOS. For this range of bandwidths converters enable low-power, high-performance
conditioning channels.
This is especially true in case a continuous-time loop filter is used as this allevi-
ates the requirements on preceding anti-aliasing filtering. In addition, most continuous-
time implementations have a better power/performance ratio than their switched-capacitor
counterparts often due to lower bandwidth requirements on the filter stages [7].
Single-bit quantization provides high-linearity. This is a major specification on the
ADC in case analog conditioning -limiting bandwidth and DR of the input signal- is traded
for digital conditioning.
Motivated by the above promise of an attractive cost/performance ratio the text con-
centrates on continuous-time, single-bit -ADCs (see also 3.2.1).
CMOS technology
The choice for a baseband mixed-signal channel justifies a further narrowing of the scope
to CMOS technology only. While CMOS is gaining ground in many application areas, it
is certainly doing so in the field of analog and mixed-signal baseband design. This follows
from the number of scientific publications in this area.
Power consumption as cost parameter

The optimization of the signal-conditioning channel is performed in a single cost dimen-
sion, being power consumption. This is certainly a viable choice for portable applications
aiming at long stand-by times and small and light-weight battery packs. In general, low
power consumption can be an important asset in view of limited heat sinking capabilities
of packages, in view of area required by fans, etc.
The analysis aims at calculating the relation between the current consumption and
the performance requirements. The maximum supply voltage is assumed to be dictated
by the technology choice.
Performance parameters
In view of the comparison of various architectures for the conditioning channel a limited
set of performance parameters needs to be identified. These parameters need to represent
a fundamental specification on a generic signal-conditioning channel, influence the bal-
ancing of analog and digital conditioning and influence the power consumption. Based
on Shannon’s theory on the capacity of a generic communication channel a meaningful
set of parameters is derived in section 2.2. These parameters relate to signal bandwidth,
signal amplitude, noise power and distortion. The associated nomenclature is discussed
in section 2.5.
6 INTRODUCTION
For now it is mentioned that only (white) thermal circuit noise is taken into account.
Other noise sources like flicker noise and shot noise only occur in a limited frequency
band and therefore are less generic.
Only differential circuits are considered such that third-order distortion dominates.
Differential operation is preferable in a mixed-signal environment anyway. Furthermore
it is assumed that all circuits operate under weakly non-linear conditions which implies
that the response at the nth harmonic is only determined by the nth order non-linearity.
1.3 Outline
The book starts with a study of a generic signal-conditioning channel in chapter 2. Apply-
ing Shannon’s theory, the choice of the performance parameters is further motivated. It is
explained how system evolution and technology advances affect the conditioning channel
and digitization is identified as a key challenge. In addition, some nomenclature is intro-
duced.
Chapter 3 presents an overview of state-of-the-art in A/D converter design and
motivates the assets of converters for digitization of the conditioning channel. For
completeness, we briefly touch upon limitations on the application of converters.
In chapter 4, power/performance relations for the building blocks of the conditioning
channel -i.e. for a major class of analog circuits, for the ADC and for the decimation
filter- are derived. This leads to conclusions on how to proceed in view of power-efficient
digitization. Further on, these results are used to compare conditioning channels, with a
varying degree of digitization, in terms of their power/performance balance.
In chapter 5, we study a full-analog and a full-digital conditioning channel. These
represent the two extremes in terms of digitization and are compared with respect to power
consumption.
Chapter 6 introduces the concept of “conditioning ADCs”. Instead of having
analog conditioning, in front of the ADC, or performing the conditioning in the digital do-
main, it is integrated into the loop. This concept is enabled by the fact that ADCs
are largely immune to interferers. The analysis of the interferer immunity and of the lim-
itations thereon, is a key topic of this chapter. In addition, various topologies are
evaluated in this perspective and a “filtering-feedback ADC” -explicitly designed for
interferer immunity- is presented. Again, the power/performance balance of the various
solutions is assessed as well.
Often, the signal-conditioning channel extends over multiple dies. In that case, dig-
itization of the conditioning channel, may lead to digitization of the inter-die interface as
well. This is the topic of chapter 7.
Chapters 8 and 9, present design examples as an illustration of the theory of chapters 5
and 6 respectively. A dual-mode receiver for FM/AM radio is considered in chapter 8. In
FM mode, the signal conditioning is highly analog. In AM mode, it is highly digitized
using multi-channel A/D conversion. In chapter 9, three implementations of a “conditio-
ning ADC” for use in a Bluetooth receiver are discussed. The first design is attractive
for systems requiring a high SNR for the digital processing. The other two designs enable
1.3. OUTLINE 7
power-efficient digitization of applications requiring less SNR. Especially the filtering-

feedback ADC is very promising because of its flexibility and an inherently low-
power architecture.
Finally, in chapter 10, general conclusions are summarized.
Chapter 2
The signal conditioning channel
This chapter describes the conditioning channel in more detail. Shannon’s theory on a
generic communication channel is applied to motivate the selection of a key set of perfor-
mance parameters. Next, the basic functionality of the conditioning channel is presented.
While the functional requirements on the conditioning channel are fixed, it is shown
that the performance targets increase due to system evolution and the digitization de-
mand. At the same time, the opportunities arising from circuit innovation or technology
evolution are limited. As such, this chapter reveals a key challenge on the implementation
of the conditioning channel in today’s playing field.
Finally, some nomenclature is introduced.
2.1 Generic communication channel
A generic communication channel is represented in fig. 2.1. A message encoded in a

signal is transported through a channel to a destination. Various noise sources add to the
wanted signal but should not corrupt the message. The capacity C of a communication
channel has been calculated by Shannon as a function of the bandwidth (BW) and the
signal/noise ratio (SNR) of the channel. In case of a Gaussian signal source and Gaussian
noise sources the following well-known formula results:
channel capacity= BW · log2 (1 + SNR) (2.1)
In order not to loose information from source to destination, a match is needed between,
on one hand, the BW and the SNR of the channel and, on the other hand, the capacity
required by the signal source and the destination. The required match is expressed by
equation 2.1.
9
10 THE SIGNAL CONDITIONING CHANNEL
Figure 2.1: A generic communication channel as described by Shannon
2.2 Performance parameters

The Shannon model is used to derive a set of performance parameters to characterize
the channel. Equation 2.1 shows that the BW and SNR of the channel are fundamental
to its performance. Of course, Shannon’s theory assumes a Gaussian signal source and
Gaussian noise sources. It has been derived for a simplified model of a communication
channel. Still, it gives an indication of the key performance parameters for an arbitrary,
non-ideal communication channel.
In a practical application, the signal source is not Gaussian. In fact, the channel
performance is often specified for a sinusoidal input signal. In addition, non-Gaussian
noise and interference sources such as distortion, aliasing due to a sampling operation
or interference from unwanted signals (for example substrate bounce) will be present.
These “noise” sources may further limit the channel capacity below what is predicted
by eq. 2.1. This limitation may be reversible for some of these ”noise”sources; for in-
stance distortion can be reduced by means of calibration. For other noise sources -like
for instance interference- this becomes more difficult. In this text, only thermal noise and
distortion are included in the calculation. Compared to substrate bounce, aliasing, etc.,
they represent a more generic specification on the implementation and very often noise
and distortion are dominant anyway.
Fig. 2.2 gives a schematic summary of this discussion: based on Shannon’s theory,
and considering practical linearity constraints, signal bandwidth, signal amplitude, noise
power and distortion are identified as the key parameters for characterizing the perfor-
mance of a generic conditioning channel. The exact metrics and the associated symbols
are discussed in section 2.5. First, the link between the Shannon channel and the signal
conditioning in an electronic system is made.
2.3 Conventional conditioning channels

The communication channel that is studied in this book is the signal-conditioning channel
in a highly digitized electronic system (fig. 1.1). In such a system the actual processing of
the information in the signal is performed in the digital domain. The signal-conditioning
2.4. EVOLUTION 11
Figure 2.2: Interference and noise sources in a communication channel
channels provide communication from the physical world to the signal processing core
of the system and vice versa. A conventional (i.e. dominantly analog) implementation of
both types of conditioning channels has been presented in fig. 1.2.
The functionality of the A/D conditioning channel includes frequency translation
from the carrier frequency to baseband. The channel needs to discriminate the wanted
signal from interferers and noise, involving selection in the frequency and in the time do-
main. The dynamic range of the input signal must be reduced to the resolution required
for digital processing and, finally, A/D conversion, including sampling and quantization,
is performed.
In addition, Shannon’s theory gives a boundary condition on correct communication
through the channel. In fact, it expresses a performance target for the channel. BW and
SNR of the conditioning channel must be high enough to provide the required channel
capacity (equation 2.1). Since the generic conditioning channel doesn’t match the ideal
Shannon channel this performance target should be interpreted in a broader sense: BW
and DR of the channel must match that of the incoming signal -with a sufficient margin-
while the SINAD and sample rate must be sufficient to accommodate the output data rate
required for the digital signal processor.
The functionality of the D/A conditioning channel is -to a large extent- complemen-
tary to that of the A/D channel. It includes D/A conversion including reconstruction and
interpolation, suppression of spurious components around multiples of the sample fre-
quency, scaling of the amplitude of the signal to the required output level and frequency
translation from baseband to the carrier frequency. A similar performance target as in the
A/D channel needs to be achieved.
2.4 Evolution
Fig. 2.3 depicts the playing field for implementing the conditioning channel. System level
requirements set the performance target on the conditioning channel. From the technology
side, opportunities and limitations on the silicon implementation of the channel arise.
Fig. 2.3 also shows the degrees-of-freedom when implementing a conditioning channel;
i.e. the definition of the architecture, the design of analog circuits and the digital signal
Figure 2.3: Playing field and degrees-of-freedom for implementing the condi-
tioning channel
processing. Below, the various elements in this picture are discussed only briefly in order
to stick to the defined scope. The discussion does indicate a future direction for the
implementation of the conditioning channel.
2.4.1 Technology advances

CMOS technology is optimized for digital processing and strives for more computational
capability per unit of area. Recently, also some attention is devoted to low-power digital
operation and to leakage problems.
On the contrary, for generations beyond the 0.25µm-node, technology evolution be-
came unfavorable for analog circuit design [8]. Here, the threats of deep-submicron seem
more numerous than the opportunities. Some advantages are available, though:
• higher gm /C, slightly higher gm /I , slightly better matching, etc.
Especially the bandwidth improvement can be exploited. However, many analog parame-
ters worsen:
• VDD drops, the output impedance of transistors decreases, the bulk effect (and the
associated non-linearity) becomes more important, thermal noise increases slightly1
1 Note that the Noise Figure of the transistor improves because its transconductance increases more than the
thermal noise.
2.4. EVOLUTION 13
Figure 2.4: System demands in terms of signal bandwidth and resolution
[9], etc.
It becomes very difficult to obtain high DC gain or linear operation for the MOS devices.
For a more thorough discussion, the reader is referred to [8], [10], [11], [12], [3], [13],
etc.
Note that the characteristics of advanced CMOS technologies match the require-
ments for ADC design. modulation exchanges resolution in amplitude for resolu-
tion in the time-domain. As such, most analog sub-blocks of the ADC are not critical
with respect to distortion, noise or matching. On the contrary, the circuit bandwidth must
increase because of the over-sampling. This is elaborated in section 3.2.2.
2.4.2 System demands

Fig. 2.4 plots a number of systems as a function of the resolution and bandwidth re-
quirement for the A/D converter. Established systems for speech, digital audio or analog
television are at the less demanding side of the diagram. More recent systems, such as the
super-audio standard, high-speed storage, GSM, etc., need more resolution and/or band-
width. Future systems, a.o. Ultra-Wide-Band data communication, etc., move further to
the extremes of the axes.
In fact, this graph demonstrates the higher bit rate demand of new systems. This
increase is due to various factors:
• more information needs to be exchanged;
• higher quality is demanded by the customer;
• more encoding “overhead” is needed because of a much denser use of the available
spectrum, space and time slots
Figure 2.5: Evolution of power consumption of ADCs for various systems [14]
and many more.

Not only are system standards becoming more demanding, also, more and more dif-
ferent systems arise. These need to co-exist or even co-operate; e.g. the fourth generation
of mobile communication systems targets seamless hand-over and multi-standard oper-
ation. The radios then should enable a wide range of air interfaces and the baseband
conditioning channel should become very flexible as well.
2.4.3 Advances in digital signal processing and analog circuit design

Digital signal processing follows Moore’s Law. This does not only translate into minia-
turization, it also results in a considerable power saving. It is calculated in chapter 4,
page 51 that the dynamic power consumption per transition reduces by a factor 1000 in
10 years time.
On the contrary, the improvement of the power/performance balance of analog cir-
cuits is much slower. In fig. 2.5, lines of constant power consumption have been added to
the graph of fig. 2.4. They indicate the power consumption in the ADC for a combined
resolution and bandwidth target. According to the International Technology Roadmap for
Semiconductors (ITRS), these lines have shifted up by a factor 10 every decade [14] over
the past.
Obviously, a huge gap exists between the pace of improvement in the power/perfor-
mance balance of analog circuits (including data converters) and that of digital processing.
2.4.4 Digitization of the architecture

For cost and flexibility reasons, the conditioning channel is being digitized. The cost
reduction aspect follows from Moore’s Law; i.e.digital blocks are subject to miniaturiza-
tion. The flexibility results from the fact that digital processing can easily be adapted to
2.4. EVOLUTION 15
Figure 2.6: Linearity and DR requirements on ADC and DAC with increased
digitization of UMTS conditioning channels (functional diagram).
new requirements, to encompass more features, to provide multi-mode and multi-standard

operation, etc.
As such, analog conditioning circuits are being replaced by digital processing. How-
ever, while reducing the number of analog blocks, the data converter shifts towards the
antenna and becomes dramatically more difficult. This is illustrated with an example
next. The example treats the digitization of the A/D and D/A conditioning channels (i.e.
receiver and transmitter respectively) for UMTS. First, the DR and BW requirements are
derived for various positions of the ADC or DAC in the channel (corresponding to vari-
ous degrees of digitization). Next, the linearity requirements are added. The calculated
figures should be considered as first-order estimates only. Their prime aim is to illustrate
the consequences of digitization on the converter specifications.
UMTS Example [15], [16], [17], [18], [19]:

UMTS is a spread-spectrum system with 5MHz-wide channels and a “chip rate” of
3.84MHz. The receive band from 2.11 to 2.17GHz is considered. Assuming a low IF
receiver, a maximum of analog filtering and VGA, then, a 3-bit, 7.68MSps ADC suffices
considering the spreading gain and the demodulation requirements. This is true both for
noise and for distortion.

Leaving out all analog VGA requires the ADC to accommodate an additional signal
range of 78dB, leading to a 16-bit ADC. Since only the wanted channel is present, the
linearity can remain as low as before.
If the analog filter is removed, the sample rate must be increased to prevent aliasing
of interferers falling within the 60MHz-wide receive band. The resolution can be lowered
by 2 bits because of the over-sampling. However, the ADC must become dramatically
more linear because of the presence of in-band blocker channels. These can be as strong
as −25dBm, while the distortion must remain below −103dBm. As such a linearity of 12
bits becomes necessary.
If the ADC is put at RF, near the antenna, then, the sample rate increases further to
about 4.4GSps. The resolution of the ADC can be lowered to 12 bits because of the very
high over-sampling.
Notice that, in the digitized receiver, the challenging linearity requirement is es-
pecially due to the presence of the interferers instead of directly relating to the wanted
signal.
For the transmitter, a zero IF architecture is suitable. Essentially, a pair of 3-bit

quadrature DACs sampled at Nyquist rate can be used to digitize the OCQPSK modu-
lated signal. However, this would require infinitely steep analog filters to suppress the
higher-order replicas of the baseband signal.
In practice, over-sampled DACs with higher resolution are used to meet the spectrum
and the spurious emission requirements. All analog filtering (except the antenna filter)
can be omitted at the expense of 16-times over-sampled, 8-bit DACs. These DACs gener-
ate quantization noise substantially below the spectrum emission limit of −49dBc /1MHz
at 12.5MHz offset. In case of straightforward multi-bit quantization, the distortion of a
single 8-bit DAC is at best −50dBc . For other encoding schemes, the quantization related
distortion can be lower. Therefore, when assuming a reasonably linear implementation,
the spectrum emission limit is respected with a margin in the quadrature configuration.
The analog VGA can be traded for digital VGA at the expense of 13-bit DACs. These
provide the 71dB of power control on top of the 3-bit resolution reserved for the synthesis
of the modulated signal while taking advantage of the 16-times over-sampling. (At low
output levels, the emission specification is set to an absolute limit and these DACs inher-
ently meet that specification.)
Direct digital synthesis -leaving out the analog mixer too- would require a sample
rate of around 4GSps and a single 11-bit DAC. (Because of the oversampling, the DAC
can have about 3 bits less than before).
Notice that the linearity requirement is set by the spurious emission limit. As such, it
equals the DR requirement.
Clearly, conventional, straightforward digitization, by pushing the ADC and DAC to-
wards the antenna, puts an enormous burden on the converters2 . In addition, new systems
2 In fact, especially for the ADC the requirements may become even stricter than what is suggested in the
example. The lack of analog gain means that the specified dynamic range must be achieved at a smaller input
2.5. NOMENCLATURE 17
Figure 2.7: A general input spectrum and definitions
demand a higher performance conditioning channel and technology evolves in an unfa-

vorable direction. The power/performance balance of ADCs improves too slowly to meet
all the demands. It is recognized that innovative channel architectures are needed to
bridge the performance gap. For instance, [20] proposes “digitally assisted analog circuit
design”; i.e. requirements on analog precision are delegated to a digital processor. Al-
ternatively, here, the focus of the architectural innovation is on optimally exploiting the
characteristics of the ADC and on “not wasting power to the interferers”. This is
clarified throughout the book.
2.5 Nomenclature
Fig. 2.7 depicts an example input spectrum to an A/D conditioning channel. This spec-
trum corresponds to a typical interference scenario in communication systems. In accor-
dance with the nomenclature in that field the notion of “wanted channel” and “interferer
channels” is used. Together, these occupy the entire “signal band”. In this book all met-
rics like BW, v̂IN , v̂MIN , v̂MAX , v 2 n,eq , DR, SNR, etc., refer to the wanted channel only, unless
stated otherwise3 . The symbols indicating a ratio, are sometimes expressed in decibels.
Then, this is mentioned as a subscript to the symbol or it becomes clear from their value.
The following metrics are indicated in fig. 2.7.
• BW: this is the bandwidth of the wanted channel. The bandwidth of the entire
band is larger by a factor p (p ≥ 1). By consequence, a maximum of p−1 interferer
channels can be present in the band.
signal than in the conventional channel.
3 For simplicity reasons, the signal at any node in the channel, is often characterized by its amplitude only.
In those cases a high impedance is assumed or only a ratio of amplitudes is important. If this is not the case, the
corresponding impedance value should be mentioned as well.
• v̂IN , v̂MIN and v̂MAX : these correspond to the instantaneous, the minimum and the
maximum amplitude of the wanted signal respectively. The amplitude levels v̂MIN
and v̂MAX are introduced to allow a discussion on the consequences of variable
gain implementation. It should be stressed that all refer to the amplitude of the
signal in the wanted channel only. In case interferer signals are present, the total
signal amplitude will be larger by a factor q (q ≥ 1) because the interferers and the
wanted signal cannot easily be distinguished from each other in the time domain.
The overall amplitude then equals q v̂IN .
• vn : this symbol is used as a short-hand notation for the rms value of an input-
referred, equivalent noise source corresponding to the wanted channel only.
• DR: the dynamic range is the ratio of the maximum signal power to the minimum
detectable signal power in the channel bandwidth. The minimum detectable power
-theoretically- approaches the minimum noise level. Then, the dynamic range also
equals the ratio of the maximum signal power to the integrated idle channel noise.
• SNR: the signal/noise ratio is the ratio of the power of the wanted signal to the
noise power4 is applied. Both are integrated over the channel bandwidth and both
are present simultaneously. This specification is set by the digital processing. For
example it can originate from a requirement for digital demodulation with a certain
bit-error rate.
In this book, it is understood that the above explained metrics refer to the wanted channel
only (unless stated otherwise).
Fig. 2.7 does not show the linearity specification. In this book, we consider inter-
modulation distortion instead of harmonic distortion (both are related and can be calcu-
lated from each other). Harmonic distortion results in components at higher frequencies,
likely outside of the bandwidth of the wanted channel. Hence, these can be suppressed
by filtering. On the contrary, intermodulation distortion may cause components inside
the bandwidth of the wanted channel. During consecutive conditioning or processing, the
distortion components cannot be discriminated anymore from the wanted signal. As an
additional argument, in many system standards, intermodulation tests are defined explic-
itly. As such, this is a meaningful distortion parameter for the conditioning channel as
well.
Assuming differential operation, we use IM3 to quantify the distortion. The IM3 -
definition is illustrated in fig. 2.8:
• IM3 : is defined for a two-tone test in which two sine wave inputs, of equal ampli-
tude and at nearby frequencies f 1 and f 2 are applied. The IM3 -value equals the
ratio of the amplitude of the intermodulation component at 2f2 −f1 (or at 2f1 −f2 )
to that of the fundamental signal at f 1 (or at f 2 ). In fact, this value depends on the
signal level. In case this signal level is not mentioned it is understood that the fun-
damental signals both have an amplitude of half the allowed full-scale amplitude
(thus at −6dBFS ).
4 For some systems, e.g. using time-division multiplexing, this definition can be adapted in an obvious man-
ner.
2.6. CONCLUSIONS 19
Figure 2.8: Definition of IM3
In contrast with the above parameters, the IM3 distortion of the various blocks is not solely
linked to the wanted signal. It depends on the overall signal amplitude consisting of both
wanted and interferer signals. Often, the interferer signals cause the dominant distortion.
In many systems (like the example of digital demodulation) distortion and noise -in a
first order approximation- have a similar effect on the processing and both must be equally
low. Then, the following relation between the IM3 -specification and the DR-specification
can be derived: √
vn 2 2
IM3 ⇔ =√ (2.2)
v̂IN /2 DR
In case values are filled out, the symbol ⇔ can be replaced by =. Also note that DR is
used and not SNR. This follows from the assumption that the major distortion problem is
due to strong interferers in the presence of weak wanted channels.
Often, the distinct contribution of noise and distortion is not important and the signal-
to-noise-and-distortion ratio (SINAD) is used. In those cases, SINAD is used when assum-
ing noise and distortion contribute equally.
2.6 Conclusions
From Shannon’s theorem, and considering practical constraints on analog circuit design,
noise, distortion and bandwidth are identified as key parameters for evaluating the chan-
nel performance for a given input signal.
Due to demanding new systems and the evolution of CMOS technology, the imple-
mentation of the conditioning channel becomes very challenging.
Digitization of the conditioning channel, further challenges the performance of the
ADC.
From the past, the power/performance balance in analog circuits and ADCs improves
by a factor of 10 per decade. In the same period of time, the power/performance balance
in digital processing, improves by a factor of 1000.
Architectural innovation is identified as a degree-of-freedom to meet the require-
ments on the conditioning channel.
In highly-digitized conditioning channels for wireless communication systems, the
ADC specifications become increasingly determined by the presence of the interferers in-
stead of only relating to the wanted channel.
Chapter 3
A/D conversion
This chapter starts with a historical overview of A/D conversion. Next, the state-
of-the-art in converter design is briefly discussed. This discussion includes some
architectural choices and circuit aspects of a design. It provides a common ground
for use in later chapters. In addition, it is motivated that a single-bit, continuous-time
implementation of a ADC with a feed forward type of loop filter, is an inherently low-
power topology.
Considering the discussion in chapter 2, this type of ADC has the prospect of
becoming an enabling block in future conditioning channels. This observation is a basis
for most of the book.
Finally, for completeness, some limitations on the use of ADCs are summarized.
3.1 Historical overview

A first patent related to delta modulation was filed in 1948 [21]. In 1952, the delta mod-
ulator was first published by de Jager and Greefkes at the Philips Research Laboratories
(fig. 3.1, [22]). The invention was inspired by the operation of the human brain: phys-
iological signals are translated into a series of electrical pulses in the nerve system as a
means for data transmission to the brain. In a similar way, pulse density modulation was
used for robust data transmission in telephony. In the single-bit, delta-modulated code all
bits are of equal weight and any bit-flip causes only a small error. In a multi-level PCM
code though a bit-flip of the Most-Significant-Bit results in a major error. The paper by de
Jager in 1952 [22], on delta modulators was the first in a massive series of delta-sigma and
sigma-delta papers. In 1960, the delta-sigma modulator was patented by Cutler [23]. In-
ose et al. [24] proposed to shift the loop filter in the forward path of the modulator in 1962.
In the 1980’s sigma-delta conversion became popular in both the A/D and the D/A part of
audio channels. In addition, instrumentation applications widely adopted converters.
By that time, a lot of theoretical work had been published, a.o by J. Candy.
In 1993, for the first time a separate session at the ISSCC was devoted to A/D
conversion. The target applications soon evolved including digital radio and 2G com-
21
22 A/D CONVERSION
Figure 3.1: The delta modulator as first published by de Jager and Greefkes in
1952
munication standards. From 2000 on, designs for 3G communications and for wireless
connectivity in the 2.5GHz band have been widely published. Anno 2005, also wire-
less connectivity in the 5GHz band and advanced wired communication standards -like
ADSL- are targeted.
3.2 State-of-the-art in A/D conversion

This section gives an overview of some architectural and implementation related aspects
of the design. As such, it provides a common ground for later reference. For a thor-
ough analysis of data converters the reader is referred to a.o. [25], [26], etc.
A formula for the signal/quantization-noise ratio (SQNR) of a modulator is re-
peated here because it gives an overview of the various architectural parameters and their
effect on SQNR:

3 2L + 1 N 2
SQNR = 2 − 1 m2L+1 (3.1)
2 π 2L
where L = order of the loop filter

N = number of bits in the quantizer
actual sample rate
m= the over-sampling factor; i.e. m = Nyquist sample rate f s
3.2. STATE-OF-THE-ART IN A/D CONVERSION 23
Figure 3.2: Block diagram of a ADC and architectural parameters
These parameters are further clarified in the block diagram of fig. 3.2. They provide the
framework for the discussion on the architectural design in the next section.
It should be remarked that eq. 3.1 results from theoretical calculations. It does not
take into account stability issues. For this reason, the achievable SQNR may, in practice,
be lower than what is predicted from eq. 3.1.
3.2.1 Architectural considerations

Normally, a ADC is dimensioned such that its quantization noise (within the targeted
bandwidth) is lower than its circuit noise. The circuit noise provides dithering and reduces
the appearance of tones. In addition, this dimensioning also yields lower overall power
consumption for the ADC. While a reduction of the noise power of an analog circuit
requires a linear increase in power consumption, the quantization noise can be lowered
by more power-effective, architectural measures. These architectural degrees-of-freedom
are discussed next.
Over-sampling factor
In an industrial realization, the value of the over-sampling factor is often dictated by the
system. For instance, the sample rate may be standardized or a specific clock frequency
is available and should be re-used. (On the contrary, the order of the loop filter and
the choice of a filter topology, as well as the number of bits in the quantizer are mostly
degrees-of-freedom in the design of the ADC.) The choice of the over-sampling
factor has a strong impact on the SQNR. Next to appearing explicitly in eq. 3.1, also L is
a function of m. This is discussed below.
Loop filter
The architectural design of the loop filter includes the choice of the filter order, the choice
between a switched capacitor or a continuous-time implementation and the choice of filter
topology.
Order of the loop filter: For a fixed over-sampling factor, the highest sensible filter
order should be implemented, because:
24 A/D CONVERSION
Figure 3.3: Continuous-time ADC with feed forward (a) and feedback (b)
loop filter
• the higher-order filter sections provide additional shaping of the quantization noise;
• still, the power consumption of these sections can be very low: since the filter is
put inside of an overall feedback loop, the noise and distortion of these stages is
suppressed by the gain of the preceding sections.
On the other hand, stability issues put an upper boundary on the filter order.
Switched capacitor versus continuous-time implementation: Key differences be-

tween switched capacitor and continuous-time implementations are discussed in [27]. A
switched capacitor implementation is more robust to clock jitter and processing spread
and its coefficients can be derived from those of a digital modulator in a straightforward
way. A continuous-time implementation often performs better with respect to power con-
sumption. In addition, it features implicit anti-aliasing filtering.
Feedback or feed forward loop filter: Fig. 3.3.a and 3.3.b respectively depict a
ADC with feedback and with feed forward compensation (for high-frequency sta-
bility of the loop). Both types of loop filter provide the same noise-shaping but a different
closed-loop transfer. (In fact, a mixture of feed forward or feedback paths can be imple-
mented as well.) In case of feedback compensation all filter stages need to linearly inte-
grate the entire output signal of the DAC. In case of feed forward compensation, the input
signal to the higher order filter stages has been filtered by the preceding stages relaxing
the linearity and bandwidth requirements. Therefore, the feed forward implementation
consumes less power. These differences are further discussed in 6.3. As a short-hand
notation these topologies are referred to as “feed forward ADC” and “feedback ADC”.
Number-of-bits in the quantizer

Multi-bit quantization provides a means for increasing resolution without increasing the
filter order or sample rate. This translates in a higher maximum stable-input range, lower
quantization noise and a better defined quantizer gain than for a single-bit design. Hence,
stability is achieved more easily in multi-bit modulators.
Single-bit quantization has the advantage of inherent linearity while conventional
multi-bit designs need ”dynamic element matching”(DEM) of the DAC. The implemen-
tation of the DEM-algorithm, of the multiple comparators and of the DAC represents a
complexity -almost- quadratic with the number of bits. Hence, the area of a multi-bit de-
sign is larger in those cases. Recently, alternative multi-bit designs have been presented:
[28] uses a multi-bit semi-digital filtering DAC to alleviate the linearity problem, in [29]
the output of the DAC is truncated in order to avoid a DEM algorithm.
The comparison of the power consumption is somewhat less straightforward: circuit
noise requirements are the same for the multi-bit and single-bit design, the hardware com-
plexity of the multi-bit design is largest but the active circuits in the single-bit design need
a higher bandwidth (because of the higher sample rate). Fortunately, the bandwidth in-
crease in the single-bit case is moderate: for a same SQNR a single-bit design needs only
a slightly higher sample rate than a multi-bit design (eq. 3.1). This implies a proportional
-thus moderate- increase in bandwidth and current consumption for the active stages. Bal-
ancing this with the complexity in the multi-bit design single-bit sigma-delta ADCs are
likely to achieve a better power efficiency for most applications. In fact, this is in line with
the discussion on the Shannon theorem and power efficiency of analog circuits (page 29):
resolution in amplitude (e.g. multi-bit quantization) leads to a higher power consumption
than resolution in time (e.g. single-bit quantization at large over-sampling). Especially the
DAC of the multi-bit design constitutes a significant part of the power consumption be-
cause noise and distortion of the DAC are directly present at the output of the ADC.
In a summary, multi-bit modulators are easier to design in view of stability while
single-bit designs outperform in terms of power/performance for most applications. In
section 3.4, it is discussed how the performance of single-bit conversion degrades at
high sample rates due to various non-linear effects (jitter, inter-symbol-interference, etc).
At these sample rates (typically required for applications with a large signal bandwidth
and a high DR specification) multi-bit designs become more attractive. At an equal target
performance, multi-bit modulators can be sampled at a -moderately- lower frequency (see
above). In addition, multi-bit quantization reduces the influence of the mentioned effects
on the overall ADC performance.
Examples
Table 3.1 compares power and performance of a few sigma-delta designs for single-bit
and multi-bit implementations1 . All achieve high resolution. The power efficiency of the
1 Only designs that are among the best published have been included in this comparison. In addition, the
pairwise comparison is between designs in the same technology generation.
26 A/D CONVERSION
Table 3.1: Comparison of single-bit and multi-bit ADCs for a few conver-
sion bandwidths
Conversion bandwidth of ∼ 200kHz
Design DR SINAD P FOM (eq. 4.4)
1-bit: [30] 92dB 90dB 8mW 3 × 10−17 J

4-bit: [31] 75dB 64dB 1.75mW 3 × 10−15 J
Conversion bandwidth of 1MHz
1-bit: [32] 76dB 75.5dB 4.4mW 1 × 10−16 J

4-bit: [33] 68dB 64dB 2.2mW 1 × 10−15 J
Conversion bandwidth of 15MHz
4-bit:[34] 67.5dB 63.7dB 70mW 2 × 10−15 J
designs is compared using a figure-of-merit (FOM) that is discussed in section 4.2. This
FOM is the ratio of the power consumption of the design to the product of SINAD and
BW. Hence, the lower the FOM the better the design. For a conversion bandwidth lower
than about 2MHz single-bit sigma-delta conversion achieves a better power/performance
ratio. For conversion bandwidths of 15MHz or higher only multi-bit designs (often in
combination with a MASH architecture) have been published.
Summarizing the above, ADCs with a continuous-time loop filter built on a feed
forward topology and using single-bit quantization, constitute an inherently low-power
architecture.
3.2.2 Implementation aspects

For later reference, some considerations on the requirements on sub-blocks of the
ADC, and some possible circuit topologies are briefly discussed here. The block
diagram of fig. 3.4 shows an example implementation of a single bit, continuous-time
ADC with second-order, feed forward loop filter. This block schematic is a popular
starting point for many continuous-time ADCs (a.o. [7], [35], [36], [37], [38] and
[39]) as well as for the designs presented in this book.
Figure 3.4: Continuous-time ADC with feed forward loop filter
V/I conversion in the input and in the feedback path

In fig. 3.4 resistors perform the linear V/I conversion in the input and in the feedback
path. In [7] operational transconductance amplifiers (OTAs) are used. These can achieve
a lower phase-shift compared to the resistors that act as a lumped RC-network. The OTAs
are less linear though.
Loop filter
In case of resistive V/I conversion the input and the feedback current are summed at the
virtual ground node provided by an OTA in an integrating, negative feedback configura-
tion. This OTA constitutes the first integrator. In case OTAs are used for the V/I con-
version the currents can be summed on the first capacitor. For both implementations the
major requirements on the first stage relate to noise and distortion The second integrator
-and all following integrators in case of a higher-order filter- could be based on a similar
stage or alternatively it consists of a gm C-section (as depicted). In the latter case, degen-
eration may be applied for linearity reasons. Parasitic poles of the integrators should be at
high frequencies in order not to add phase delay to the loop. Therefore, OTAs are widely
used. In a mixed-signal environment the OTA is normally based on a differential pair. In
addition, the input transistors can be cascoded for linearity reasons. Possible topologies
and their specific advantages are:
• OTA with telescopic cascode: requires only a minimum number of current branches;
• OTA with folded cascode: allows a large output swing for the integration.
In the latter topology the power consumption is doubled because of the double amount of
current branches. Therefore, the designs presented in this book use a telescopic cascode
topology. Still, if the supply voltage of future CMOS technologies continues dropping it
may become necessary to use a folded cascode topology instead.
Other topologies based on a differential pair include:
28 A/D CONVERSION
• differential gain stage followed by source-follower: provides a low output im-

pedance for fast settling;
• a cascade of two differential gain stages with Miller compensation: provides high
DC-gain.
These topologies have an additional bandwidth limitation on the internal node between
the consecutive stages. Therefore, they are not preferred for A/D conversion at a high
sample rate.
Quantizer and DAC

The quantizer is not critical due to the loop operation. It should not add significant delay
though. The feedback DAC consists of switches connecting the resistors to the negative
or the positive reference voltage depending on the polarity of the output data.
Further detail on possible implementations can be found in a.o. [40]. Notice that the
proposed implementation allows technology scaling and operation at a low supply volt-
age: only the input stage is critical. It needs to have a reasonable output swing without
going into saturation. Fortunately, its input is a virtual ground node and therefore the
input signal swing is negligible.
3.2.3 Performance metrics for ADCs

The performance of a ADC with respect to noise and stability is typically character-
ized in a diagram as depicted in fig. 3.5. The SNR increases with the input level until the
peak-SNR is reached. Beyond this input level the quantization noise rises and the SNR de-
grades because of a reduced stability. The mechanism behind this large-signal instability
is the fact that the DAC cannot feed back a large enough signal in time to compensate for
the input signal (i.e. the phase and the gain margin of the loop become too small). The
error signal and the internal signals in the loop (i.e. the outputs of the various integrators)
therefore grow and further reduce the gain and the phase margin resulting in instability.
For a further discussion on stability the reader is referred to [25], [40].
In this book, the input level corresponding to the peak-SNR is defined as the maxi-
mum stable input. The DR is defined relative to this input level. The designs presented in
this book (chapter 8) typically achieve a maximum stable input corresponding to a mod-
ulation depth of the digital output of ∼ 70%2 . This is 3dB below digital full-scale. This
modulation depth corresponds to a good compromise between aggressive noise shaping
and stability. Designs achieving a higher modulation depth have a lower peak-SNR. Vice
versa, designs with a higher peak-SNR only allow a smaller modulation depth3 .
2 The electrical value of the maximum stable input is implementation specific: next to the allowed digital
modulation depth it also depends on the output current/voltage of the DAC in the feedback path.
3 Notice, in some texts the definition of DR is based on a -theoretical- 100% modulation depth compared to
the idling noise. Especially for designs with a poor stability this definition results in misleading, high values for
both DR and peak-SNR.
3.3. ADCS IN FUTURE CONDITIONING CHANNELS 29
Figure 3.5: Definition of some performance metrics for ADCs
3.3 ADCs in future conditioning channels

Considering the discussion in chapter 2 on the playing field for future conditioning chan-
nels, it is motivated that ADCs can be a key enabler for these channels. It is shown
that ADCs are built on an inherently low-power architecture and that this is due to two
main characteristics; i.e. over-sampling is used and overall feedback is applied. The first
bullet relates to the fact that modulators trade resolution in amplitude, i.e. SINAD, for
resolution in time, i.e. BW. It is mentioned in section 2.4.1, that for this reason, based
conditioning channels seem attractive for integration in advanced CMOS technologies.
Here, it is added that, in view of Shannon’s theorem and in view of power consumption
in analog circuits, this is an asset leading to a low-power solution.
The second bullet is important for a comparison with Nyquist A/D converters. The
overall feedback relaxes various accuracy requirements on analog sub-blocks.
Finally, the previous, rather intuitive reasonings are supported by a survey of pub-
lished power/performance of A/D converters.
3.3.1 The Shannon theorem and based signal conditioning

The Shannon theorem considers the data rate through a channel. Here, it is applied in
a more generalized way; i.e. on an electrical signal. This signal -just as the Shannon
channel- is characterized by a bandwidth and a SNR. These parameters determine the
data rate that the signal can contain. For sure, the signal properties may not fully meet
the Shannon assumptions (a.o. Gaussian signal and noise sources), but this discussion is
30 A/D CONVERSION
meant as an intuitive reasoning only.

Hence, applying Shannon’s theorem to a true-life signal:
data rate ∼ BW
data rate ∼ log SNR
where data rate, BW and SNR relate to the signal. This means that, for a fixed data rate,
resolution in amplitude, i.e. SNR, can be exchanged for resolution in time, i.e. bandwidth.
This property is the basis of encoding.
Taking into account the following considerations on power consumption in analog
circuits:
power consumption ∼ BW
power consumption ∼ SNR
assuming the latter is dominated by thermal noise. Here, power consumption, BW and
SNR relate to the circuit.
The bandwidth and SNR requirements on the analog circuit are determined by the
bandwidth and the SNR of the signal. This links the power consumption of the analog
circuits in the conditioning channel to the encoding of signals. Hence, it can be con-
cluded that:
• in view of power-efficiency of the analog conditioning circuits, an encoding scheme

that trades resolution in time for resolution in amplitude, should be used.
This is exactly what happens in a modulator.
3.3.2 Comparison of Nyquist and based signal conditioning

A conditioning channel consisting of an anti-alias filter and a Nyquist A/D converter is
depicted in fig. 3.6.a. In fig. 3.6.b, the block diagram of a conditioning channel with
A/D conversion is shown. Because of the over-sampling, the ADC reduces the
requirements on the anti-alias filter. In fact, it is analyzed in chapter 5, and demonstrated
with examples in chapters 8 and 9, that in many based conditioning channels the
anti-alias filter can be omitted. As such, it is not included in the present analysis either.
The loop filter and the quantizer of the ADC, can be compared to the anti-alias
filter and to the Nyquist ADC, respectively.
The specifications on the anti-alias filter in the Nyquist channel are challenging:
• all stages of the filter -or generalized, all stages of the Nyquist channel- need to
achieve the full dynamic range requirement4 . Even though the consecutive stages
can have a higher noise (inversely proportional to the preceding gain), they must
remain linear over a larger input range, as well;
4 For simplicity reasons, variable gain amplification is not considered in this comparison, even though, it
would reduce the dynamic range requirements for following blocks. It is shown in chapter 6 and 9 that variable
gain can also be implemented in ADCs, yielding a similar benefit.
3.3. ADCS IN FUTURE CONDITIONING CHANNELS 31
Figure 3.6: Comparison of an analog conditioning channel with Nyquist ADC

(a) to a single-bit, based conditioning channel (b)
• the accuracy requirements on the anti-alias filter in the Nyquist channel can be very
stringent. These requirements relate to the pass-band ripple, the transition band and
the suppression required at half the sample rate.
On the contrary, the design of the loop filter for the ADC is much more relaxed:
• the consecutive filter stages of the loop filter can have increasing noise and dis-
tortion. Because of the overall feedback, these contributions are suppressed towards
the output;
• the accuracy of the loop filter transfer is not too critical. The gain of the loop filter
needs to be ”high”within the conversion bandwidth, while the accuracy of the gain
is not important. A deviation on the cut-off frequencies can be taken into account
when defining the gain and phase margin of the loop.
For completeness, it is mentioned that the DAC in the feedback path of the modulator
is a critical block. Still, for a single-bit design the required accuracy is normally easily
achieved. On the contrary, the requirements on the Nyquist ADC are strict:
• the number-of-bits and the accuracy of the sub-blocks of the ADC directly affect
the resolution in the channel;
• theoretically, the Nyquist ADC can be sampled at a frequency equaling twice the
signal bandwidth; i.e. fs = 2BW. In practice, fs > 2BW is required in order to relax
the transition-band for the anti-alias filter and in order to reduce the amplitude and
phase error of the sampling action.
The quantizer in the ADC can be inaccurate:

32 A/D CONVERSION
• the number-of-bits in the quantizer can be lowered (ultimately to a single bit) be-
cause the quantization noise added in the quantizer is shaped by the transfer of the
preceding loop filter. Similarly, all other noise and distortion sources in the quan-
tizer are not critical either;
• the sample rate is higher, but since CMOS-technology is optimized for high-speed
switching, this should not be a problem for most applications (see discussion in
section 3.4.
In order to make a completely fair comparison, the power consumption of the decima-
tion filter in fig. 3.6.b should be taken into account as well (see section 4.5). The above
comparison, does indicate that the use of A/D conversion significantly relaxes the
accuracy requirements for most analog blocks in the channel. This is confirmed by the
comparison of published power/performance for both types of A/D converters, below.
3.3.3 Survey of published power/performance values

The power/performance balance of various converters can be benchmarked using the
Figure-of-Merit (FOM) according to eq. 4.6: the lower the FOM-value, the better the
ADC. This FOM is defined and motivated in section 4.2. Here, it is used for a first-order
comparison of the power/performance balance of ADCs that may differ both in architec-
ture, implementation style, technology, etc.
The data for this analysis is taken from [41], evaluating this FOM for Nyquist ADCs,
and from table A.1 in appendix A, listing this FOM for discrete-time and continuous-time
ADCs. From this data, the following is observed:
• for Nyquist ADCs, the best FOMs equal ∼ 1pJ or higher;
• for discrete-time ADCs, FOMs of 0.35pJ and 0.5pJ have been reported ([42]
and [43] resp.), but most exceed 0.9pJ;
• for continuous-time ADCs, a lot of designs achieve a FOM in the range of
0.3pJ to 0.6pJ .
This comparison favors continuous-time A/D conversion5 and confirms the state-
ments of sections 3.3.1 and 3.3.2. In fact, the lowest FOM is achieved for single-bit feed
forward implementations. This was anticipated in section 3.2.1.
3.4 Limitations of A/D conversion

From the above, ADCs seem to enable low-power, high performance conditioning
channels that are scalable to deep-submicron CMOS technologies. Historically, the ap-
5 For a fair comparison, the power consumption in the decimation filter of the designs should be taken
into account as well. However, technology scaling is very helpful in reducing this contribution.
3.4. LIMITATIONS OF A/D CONVERSION 33
plication area of converters has been restricted to low bandwidth, high resolution
systems. Even in advanced technologies, though, the achievable sample rate is hamper-
ing the adoption of conversion for high-performance, wide-band applications. In the
discussion, linear and non-linear limitations are distinguished in the sense that linear lim-
itations (such as phase margin) do scale with technology and current consumption, while
non-linear limitations must be overcome by a more than proportional power increase or
by innovative circuit design.
3.4.1 Linear limitations

The sample clock in a ADC -by definition- is significantly higher than the signal
bandwidth. It sets the time interval in which the feedback pulse needs to be processed by
the loop filter and applied to the comparator, timely for the next sample moment. Hence,
the sample frequency sets the requirement on the circuit bandwidth. This is different from
an entirely analog feedback loop where the bandwidth requirement on the circuits is rel-
ative to the signal bandwidth. Some important practical bandwidth limitations include
the limited unity-gain-bandwidth of the analog circuits, delay in the feedback path (time-
constants of wiring, resistors, etc.), delay in the quantizer, etc. These limitations reduce
the phase margin of the loop. Consequences are a decreased stability (a lower maximum
input signal) and a reduced noise-shaping (an increase of the in-band noise), resulting in
a smaller dynamic range. Some of the listed limitations can be solved by increasing the
quiescent current. In addition, various methods have been published to compensate for
the parasitic delay. These methods use pulse positioning of the DAC or add compensating
zero-ing paths in the loop filter ([44], [45]).
Ultimately, a fundamental limitation on the sample rate of the loop is due
to the f T , f max or any other frequency characterizing the maximum operation of the
transistors[46].
A survey of papers at the ISSCC conferences from 1999 until 2003 gives an
impression of state-of-the-art sample-rates for ADCs in a few generations of CMOS
technologies (see table 3.2). It also lists a brief description of the architecture. Ap-
parently, sample rates are limited to a few 100MHz (in CMOS, for high performance and
robust designs) and do not really increase in advanced technologies. The reason for this
discrepancy should be looked for in additional, non-linear limitations.
3.4.2 Non-linear limitations

As the sample rate of the ADC increases the pulse duration of the feedback DAC
drops and some parasitic effects become more important: jitter on the DAC clock, inter-
symbol-interference in the DAC, etc. The impact of these effects is strongly linked to
the actual implementation of the DAC. Especially single-bit switched current DACs using
square wave output signals are prone to clock jitter. DACs with reduced jitter sensitiv-
ity are presented in [50], [51] and [30] respectively using a cosine wave and a switched
capacitor discharge pulse for the DAC. Inter-symbol-interference problems can be solved
34 A/D CONVERSION
Table 3.2: State-of-the-art sample rate for ADCs in various CMOS tech-
nologies
Reference Technology Sample rate Architecture
[6] 0.13µm CMOS 160MHz Mash 2-2, DTa

[37] 0.18µm CMOS 153MHz Low-pass, CTb
[47] 0.25µm CMOS 185MHz Low-pass, DT
[48] 0.35µm CMOS 400MHz Band-pass, CT + DT
[49] 0.6µm CMOS 100MHz Low-pass, CT
a DT=Discrete Time
b CT=Continuous Time
by differential operation and return-to-zero schemes. [52] and [28] present DACs with
increased robustness to both inter-symbol-interference and jitter.
One other sample rate related problem in a ADC may be due to a limited slew-
rate of the integrators. This results in non-linear, incomplete settling and therefore dis-
tortion [53]. However, slewing is easily prevented in most continuous-time designs since
the peak currents are moderate compared to the quiescent current. For example, in the
designs presented in chapter 9, high-frequency current transients are less than 160µA
(peak-to-peak and differential) for the first integrator while its quiescent current is as high
as 500µA6 .
As more (and different) bandwidth limitations become important the current con-
sumption may show a more than linear growth. Innovative circuit design is needed to
overcome these limitations and to further extend the application area of ADCs to
higher sample rates and higher resolution. A more elaborate study of these limitations is
outside of the present scope.
Finally, the adoption of A/D converters, is also hampered by the fact that there
exists no solid proof of their stability. Similarly, some aspects of the converter remain
difficult to analyze. Again, this is outside of the present scope.
3.5 Conclusions
Next to providing a common ground for later reference, the discussion in this chapter
leads to the conclusions listed below.
6 In switched-capacitor designs, slewing is more likely to occur due to the large charge-transfer transient
currents. However, the resulting distortion can be kept low if there is sufficient time to settle to the correct final
value [54].
3.5. CONCLUSIONS 35
ADCs with a continuous-time loop filter, built on a feed forward topology and
using single-bit quantization, constitute an inherently low-power architecture.
The requirements for ADC circuit design, match the characteristics of advanced
CMOS technologies; i.e. most circuits are nor critical with respect to distortion, noise or
matching but they do need a high bandwidth.
The basic reasons for the excellent power/performance of ADCs are the use of
over-sampling and the use of overall feedback.
In a survey of published ADCs, the power/performance balance of continuous-time
ADCs outperforms that of other classes of ADCs.
The application area of A/D conversion is limited by sample rate related prob-
lems. At present, sample rates of a few 100MHz are feasible in standard CMOS.
Chapter 4
Power consumption in channel

building blocks
The aim of this chapter is in analyzing the power consumption of the based con-
ditioning channel as a function of the performance parameters. First, a brief review of
literature on this topic is presented and some commonly used “figures-of-merit” are in-
troduced. Next, a dedicated analysis of the power consumption in the major building
blocks of the channel is conducted. A power/performance relation is derived for the ana-
log circuits, for the ADC and for the decimation filter separately. The comparison
of the different power/performance relations yields an understanding of possible trade-
offs when exchanging analog for digital conditioning. As such, these results are used for
the analysis and comparison of various channel architectures in chapters 5 and 6. The
power/performance analysis of the ADC supports the use of one particular figure-of-merit
-and a small correction thereon- throughout the book.
4.1 Literature on power/performance analysis

In literature, many articles on power consumption versus performance are found. Calcu-
lations on fundamental limits for the power consumption of a single transistor or a basic
cell can be found in a.o. the work of Vittoz [55] (calculating the power consumption per
pole as a function of dynamic range for analog) and the work of Meindl [56] (giving the
power consumption for a single binary transition).
Other references present similar calculations on basic circuits. [57] calculates the
minimum power consumption of a continuous-time A/D converter limited by the
noise of the input and the feedback resistor. [5] analyzes the absolute minimum power
consumption of analog circuits as a function of a specific SINAD over the full signal band-
width. In [8], [58] the model is refined by including the circuit’s intrinsic distortion due
to a finite output impedance.
37
38 POWER CONSUMPTION IN CHANNEL BUILDING BLOCKS
Many power/performance relations and extrapolated predictions on circuit perfor-

mance have been published. [1], [4], [59], [12] and many other references treat power
requirements due to limitations on matching, bandwidth, dynamic range, etc.
Most of the above referenced publications are not adequate to conduct the power/per-
formance analysis of the conditioning channel because they concentrate on a single tran-
sistor. Others do not include distortion. On the contrary, [5] does include both noise and
distortion and therefore could be used to analyze the power/performance relation of the
analog part of the conditioning channel. Still, a dedicated analysis yielding simplified
results is presented in sections 4.3, 4.4 and 4.5. It allows straightforward comparison with
the analysis on the ADC. In addition, it provides a physical understanding and detail-
ing of the “figures-of-merit” that are introduced next.
4.2 Figures-of-merit
Power/performance relations have been studied on a fundamental level in literature. More
empirical, but very practical for comparison of the power/performance ratio achieved
in different designs are the so-called FOMs. FOM stands for figure-of-merit. A FOM
does not necessarily express a scientific relation between the power and the performance
parameters. Rather, it is a means for straightforward comparison of circuit performance
by combining a few important specifications in a single number. Some popular FOMs are
discussed below.
4.2.1 FOM related to thermal noise

A manifold of variations exist on a FOM that combines power consumption, dynamic
range and bandwidth [60]. Often used is:
P
FOMDR ≡ (4.1)
DR BW
This FOM follows from thermodynamics. It is typically used for analog circuits where
the power consumption is inversely proportional to the noise power. Likewise, it is very
applicable to A/D converters because these are normally designed such that thermal
noise of the analog input stages dominates over the quantization noise (section 3.2).
The lower limit on this FOM is calculated for a passive network consisting of a single
resistor R. Then:
v̂IN2
DR = (4.2)
8kTRBW
and the theoretical minimum power consumption equals:
v̂IN2
P= (4.3)
2R
such that the FOM equals 4kT, i.e. 1.6 × 10−20 J. This number is a lower limit and can be
used for benchmarking the FOM of active circuits.
4.2. FIGURES-OF-MERIT 39
4.2.2 FOM including distortion

Often, SINAD is used to take distortion and spurious components into account:
P
FOMSINAD ≡ (4.4)
SINAD BW
where SINAD is expressed as a ratio of powers. This FOM is justified by a derivation
further on in this chapter that leads to eq. 4.10 for analog circuits and to eq. 4.16 for
ADCs. As such, this FOM can be used for comparison of both types of circuits.
4.2.3 FOM related to signal resolution

For Nyquist A/D converters a FOM related to the number of quantization levels is com-
monly used. It is expressed in terms of the effective number of bits (ENOB) instead of
SINAD:
SINADin dB − 1.76
ENOB = (4.5)
6.02
Furthermore, the minimum of twice the effective resolution bandwidth f E R B and the
Nyquist sample rate is used instead of the signal bandwidth. Hence:
P
FOMENOB ≡ (4.6)
2ENOB min(2fERB , fs )
The key difference between the latter and the previous FOMs is in the fact that the power
consumption is compared to signal resolution (≈ 2ENOB ) instead of the ratio of signal
power over noise and distortion power (≈ 22ENOB ). The FOM of eq. 4.6 may originate
from the fact that for Nyquist A/D converters the hardware complexity (i.e. number of
comparators, gain stages, etc.) is proportional to the number of quantization steps [41].
On the contrary, this is not true for a ADC: the relation between the hardware com-
plexity of a ADC and the number of quantization steps is not straightforward. The
increase in complexity -and thus power consumption- is a complicated function of the
filter order, the over-sample ratio and the number of quantization steps. The effect of the
various parameters on ENOB is expressed in equation 3.1. The quantization noise power
can be reduced by a less-than-linear increase in hardware complexity. Moreover, A/D
converters are normally designed such that thermal noise of the analog input stages dom-
inates over the quantization noise anyway (section 3.2).
In general, this FOM can be criticized because it relates power consumption in the
numerator to resolution in amplitude (instead of signal/noise power in the denomina-
tor. Consequently, it values resolution in amplitude equal to resolution in time, which
is conflicting with the Shannon theorem. From a thermodynamics point of view, power
consumption of an ADC should be related to its noise power instead of its amplitude res-
olution.
When plotting power/bandwidth versus SINAD for a large number of published
ADCs, the FOM of eq. 4.6 closely predicts both the best-fit trendline and the line con-
necting the “best”data points; i.e. the ADCs with lowest power/bandwidth for a given
SINAD. This exercise is reported in [61] for a set of Nyquist and ADCs. Here, it is
Figure 4.1: Overview of ADCs as listed in appendix A
repeated for ADCs only. The data is listed in table A.1 of appendix A and comprises
around 70 designs published from 2000 until 2004. The resulting chart of fig. 4.1 proves
that, also for ADCs, the FOM of eq. 4.6 closely predicts the power/bandwidth versus
SINAD trendline. In [61], this anomaly between the theoretical prediction (∼ 22ENOB )
and the practical trendline (∼ 2ENOB ) is attributed to technology constraints, architectural
overhead and various implementation related limitations. Here, we add that this trend
only shows up when averaging over a lot of designs with different architectures. Within
one class of architectures, for example, single-bit ADCs with a feed forward loop
filter, a steeper slope is observed. Connecting the “best”designs in fig. 4.1 in a piece-
wise fashion, instead of using a straight line, yields some additional insights. Converters
achieving a high SINAD do follow the relation ∼ 22ENOB of eq. 4.4. Likely, thermal noise
is determining the power/performance relation of these converters. For converters with a
lower SINAD, the slope becomes less steep. Possibly, part of the power is consumed in
some architectural overhead. That part may not scale -or may scale only weakly- with
SINAD. Combining this part of the power consumption with contributions that do scale
proportionally with SINAD, may be the reason why the slope averages and the curve fol-
lows ∼ 2ENOB . Most of the converters with a SINAD below 60dB target a reasonably
wide bandwidth (see table A.1). Likely, the power consumption is then dominated by
bandwidth related limitations (see section 3.4) or by a fixed architectural overhead (e.g.
MASH converters). This would explain the flat curve.
For sure, the above explanation remains speculative to some extent. The FOM of
eq. 4.6 seems very well suited for predicting the state-of-the-art minimum power con-
sumption that is required to meet a certain performance target. Hence, it is very often
4.3. POWER CONSUMPTION IN ANALOG CONDITIONING CIRCUITS 41
used for benchmark purposes. Still, in this book, especially the FOM of eq. 4.4 is used.
As motivated, for ADCs, it is meaningful from a physical point-of-view. Therefore,
it is well-suited for extrapolation of power consumption of a specific ADC implementa-
tion to a varying performance target or vice versa. In addition, in the next section, it is
demonstrated that FOMSINAD is valid both for the ADC as well as for analog cir-
cuits. Hence, it allows comparison of the power/performance relation of both. Finally,
this FOM can also be used for benchmarking purposes. It yields the same ranking as
FOMENOB as long as ADCs of the same bandwidth are compared. This constraint is re-
spected in all benchmarks in this book. For completeness, appendix A includes a table
listing both FOMENOB and FOMSINAD for a large set of published ADCs.
4.3 Power consumption in analog conditioning circuits

A relation between current consumption on one hand and thermal noise, intermodulation
distortion and signal bandwidth on the other hand is mandatory for a basic analysis of the
channel (see chapter 1). This relation is first derived for the analog conditioning circuits.
Analog signal conditioning includes amplifying the signal, reducing the signal bandwidth
and dynamic range. Hence, filters and amplifiers (having a fixed or a variable gain) are
the main building blocks. For simplicity, the analog conditioning circuits are assumed to
be based on a differential pair. In practice, this is often the case. In addition, it is in line
with the assumption of differential pair based integrator stages in the implementation
(page 27). This analogy allows a generalized analysis of the channel and the exchange of
analog filter stages for increased A/D performance (see chapter 5). The analog part of the
conditioning channel becomes a cascade of differential pairs, each performing a specific
function (filtering or amplification). On the other hand, the above assumption excludes
specific linearization techniques like a.o. [62], [63], [64], [65]etc.
4.3.1 Power/performance relations

The analysis focuses on the topologies based on a simple differential pair, a differential
pair in a global feed-back configuration and a degenerated differential pair as depicted
in 4.2.a, b and c respectively. Of course, multi-stage designs can be analyzed in a similar
way.
For these topologies, first order relations between the current consumption and
noise, distortion and signal level are derived in appendix B. It is assumed that -from all
active devices- the input transconductors contribute most to the noise. For the calculations
on distortion only the non-linear V/I conversion is included because it is directly related
to the quiescent current. Other non-linearities are intrinsic [8]. Their effect can only be
reduced by proper transistor dimensioning or circuit design.
Figure 4.2: Topologies with differential pair
Under these conditions, the following power/performance relation is derived in ap-

pendix B. α
DR BW v̂IN
I ∼ √ (α = 0, 1 or 2/3) (4.7)
v̂IN2 IM3
with:
• α = 0 or 1 for fig. 4.2.a: this assumes the linearity requirements are relaxed such
that an open loop configuration becomes possible: the current consumption is either
fixed (α = 0) or the over-drive voltage vGT and quiescent current of the input
transistors are adapted to the specification (α = 1).
• α = 2/3 for fig. 4.2.b and c: global or local feedback is used in case of challeng-
ing linearity requirements. The input transistors are biased near weak-inversion.
Hence, the intrinsic linearity of the differential pair is small (i.e. vGT is small) but
the linearizing effect of the increased feedback gain (i.e. a large gm ) is dominant
because it appears with power 3 in eq. B.7.
The latter implies that, in view of power consumption, it is beneficial to apply feedback
instead of relying on the inherent linearity of the input transistors. Therefore, feedback
(i.e. an α-value of 2/3) is assumed from now.
4.3.2 Discussion
The above result is commented and compared to literature.
Effect of circuit bandwidth

The presence of BW in eq. 4.7 only refers to the noise bandwidth. In addition to these
equations, the quiescent current of all topologies is also proportional to the required circuit
bandwidth. Depending on the exact specifications the current required due to noise and
4.3. POWER CONSUMPTION IN ANALOG CONDITIONING CIRCUITS 43
distortion (eq. 4.7) or the current required for the circuit bandwidth (proportional to the
signal bandwidth) will be dominant.
Results in terms of SINAD

Often, circuit performance is specified in terms of SINAD (e.g. in data transmission). In
cases where distortion and noise contribute in an equal way to the SINAD value, eq. 2.2 is
valid. Hence, the following relations are assumed:
SINAD ∼ DR and SINAD ∼ IM3−2 (4.8)
Substituting these in eq. 4.7 yields:
BW SINAD1+α/4
I ∼ (4.9)
v̂IN2−α
Other applications (like audio and video) have a different tolerance on linearity compared
to noise and equation 2.2 does not hold. In the remainder, it is used though in order to
simplify the expressions.
It is mentioned here, that for various proportionality relations in this chapter, the
units of the left and right hand side differ. This is due to the omission of vGT , kT,etc. By
consequence, these relations only express a proportionality between the value at the left
and at the right hand side.
Comparison to literature
As discussed in section 4.1, most work on power/performance relations only takes into
account BW and DR [55], [12], [3] and [57]. This corresponds to the results of the α = 0
case in the analysis above.
Reference [5] does consider distortion. It presents a.o. a relation for the absolute
minimum power of an analog circuit as a function of the required SINAD and signal
bandwidth. The key assumption in the analysis is that the load of the circuit behaves ca-
pacitively over a large part of the signal bandwidth. Hence, this condition is fulfilled for
the α = 2/3 case in the analysis above because the dominant pole of the open loop transfer
function is much smaller than the signal bandwidth.
A further assumption in the referenced paper is the fact that the output swing is
rail/rail. Thus, for a fixed gain, VD D can be substituted by v̂IN . The above-derived expres-
sion for the current consumption (eq. 4.9) is multiplied with the supply voltage to obtain
the power consumption. Next, VD D is substituted by v̂IN , yielding:
BW SINAD7/6
P∼ 1/3
(4.10)
v̂IN
This is in agreement with the results in [5]. Hence, this comparison serves as a sanity
check for the calculations in this chapter.
4.4 Power consumption in a ADC

As mentioned in section 1.2 and further motivated in section 3.2, the present analysis of
power consumption concentrates on continuous-time single-bit A/D converters with
a feed-forward loop filter. This analysis is conducted for the reference design depicted in
fig. 3.4.

In [40], it is shown that, in view of an optimal power/performance balance, the ADC
should be designed such that the input resistance Rin (as defined in fig. 3.4) dominates in
the overall noise. Then:
v̂IN2 /2
DR = (4.11)
8kTRin BW
In the same reference [40], the distortion of the topology of fig. 3.4 is analyzed in
terms of HD3 and assuming the differential pair constitutes the dominant non-linearity.
From that, the intermodulation can be derived:
2
3 v̂IN 1
IM3 = (4.12)
16 vGT (gm Rin )3
with vGT the over-drive voltage of the input transistors of the first stage. Note the follow-
ing:
• the ratio of the signal swing at the input of the transconductor to vGT determines
the inherent linearity of the input transistor
• the feedback operation (expressed by the product gm Rin ) reduces this swing
The latter effect is dominant and therefore lowest distortion is achieved if the input tran-
sistor is biased near weak inversion. Finally, the following relation is derived by solving
Rin from eq. 4.11 and substituting it in eq. 4.12:
DR BW
I ∼ 4/3
(4.13)
IM3 1/3 v̂IN
The power consumption of other building blocks of the ADC has not been included
in this calculation. These blocks have been discussed in section 3.2.2. Their current con-
sumption can be neglected or it depends in the same way on the performance parameters.
This is motivated next:
• the DAC: It can be implemented using switches only. As CMOS technology has
been optimized for switching the associated power consumption is negligible.
• the quantizer: It consists of the comparator and some logic gates. Again, the power
consumption associated with the switching of a few gates can be neglected. In gen-
eral the current in the comparator is set by the requirements on the gain-bandwidth
product, noise and offset. In the ADC only the BW requirement is important
4.4. POWER CONSUMPTION IN A ADC 45
in view of parasitic loop delay. Other artifacts are less important due to the high
loop gain that precedes these error sources. Hence, the current consumption in the
quantizer is small.
• other filter sections: The higher-order filter sections are normally down-scaled
copies of the first stage. Therefore, they obey the same power/performance relation
and their contribution in current consumption -although much smaller than that of
the first stage- can be accounted for in the value of the proportionality parameter.
4.4.2 Discussion
The mathematical relations are commented, rephrased and compared to popular FOMs
for ADCs.
Effect of circuit bandwidth

The above analysis is valid as long as the current consumption is set by noise and dis-
tortion requirements. In case sample rate associated limitations (see section 3.4) become
dominant these will further increase current consumption.
Although BW explicitly appears in eq. 4.13 it actually only refers to the noise band-
width as introduced in eq. 4.11. The bandwidth requirements on the sub-circuits of the
ADC are not truly taken into account. For most ADCs, the power is indeed deter-
mined by noise and distortion instead of bandwidth requirements. Hence, eq. 4.13 is
used to extrapolate the power consumption of most continuous-time ADCs.
Results in terms of SINAD

Again, equation 4.13 can be expressed in terms of SINAD assuming noise and distortion
contribute in a fixed ratio. Eq. 4.8 is substituted in 4.13 yielding:
SINAD7/6 BW
I ∼ 4/3
(4.14)
v̂IN
This result is the same as that for the analog circuits using feedback.
Comparison with popular FOMs

In section 4.2 it was motivated that for design purposes the FOMs in terms of DR and
SINAD (eq. 4.1 and 4.4 resp.) are more meaningful for ADCs than the FOM in terms
of quantization steps (eq.4.6). These FOMs are now compared with the derived analytical
relations.
First, the popular FOM of eq. 4.1 (including DR only) is considered. It is compared
to the following modified FOM derived from eq. 4.13:
4/3 1/3
P v̂IN IM3
FOM = (4.15)
DR BW VD D
It expresses the same linear relation between power consumption of the ADC and the
product of bandwidth and dynamic range. In addition, it shows the way to including
the effect of intermodulation distortion and the effect of v̂IN and the supply voltage VD D .
Notice, the dimension of this FOM deviates from that of all other FOMs. This FOM is
used a.o. in chapter 8 to compare the performance of an analog VGA to that of a ADC.
Secondly, eq. 4.14 is used to motivate the use of FOMSINAD (eq. 4.4) throughout the
book. The analytically derived eq. 4.14 yields the following modification as compared to
eq. 4.4:
4/3
P v̂IN
FOM = (4.16)
SINAD7/6 BW VD D
The modification compared to the conventional FOMSINAD is three-fold:
• v̂IN is taken into account: A fixed DR is achieved at less power as the input signal
becomes larger: power decreases quadratically with v̂IN . The opposite is true for
distortion: for a fixed IM3 -value gm must increase almost proportionally with v̂IN
(eq. 4.12). The combined effect results in an -almost- linear relation between the
FOM and v̂IN .
• the FOM is normalized to the supply voltage: In fact, this eliminates VD D from
the FOM leaving the quiescent current -instead of the power consumption- in the
numerator. Indeed, in the above analysis performance was assumed independent of
VD D and only related to the bias current1 .
• SINAD is included more accurately: because of distortion, SINAD increases more
than proportionally with power consumption.
Furthermore, the following remarks on the application of the modified FOMs hold:
• modified FOMs and technology scaling: The normalization to VD D eliminates the
dominant effect of technology scaling on the FOM. Therefore, these FOMs espe-
cially express the cleverness in exploiting the technology or in dealing with it. Still,
smaller effects of technology scaling, like the increasing gm /I or the decreasing
inherent linearity of the transistor, remain included.
• modified FOMs and different topologies: Notice that the same (simplified)
FOMs would have been derived for any topology different from fig. 3.4 as long as
a differential pair constitutes the dominant power consumption.
It would have been interesting to compare some published ADCs according to FOMDR
and the suggested modification thereof. Unfortunately, many publications do not mention
the input voltage. Therefore, important references could not be included and the com-
parison becomes less meaningful. Still, this exercise is conducted for a few designs
published by Philips Research for which all required data is internally available. Table 4.1
ranks the publications according to the modified FOM (eq. 4.15) in the utmost right col-
umn. From this table a few observations are made:
1 This assumption is valid as long as the ADC is easily scalable. This seems to be the case for the implemen-
tations in chapter 8 for the next CMOS generations to come.
4.5. POWER CONSUMPTION IN DIGITAL CONDITIONING CIRCUITS 47
Table 4.1: Comparison of conventional and modified DR-based FOM
Reference Technology FOMDR FOMDR,modif

eq. 4.1 eq. 4.15
[30] CDMA 0.18µm CMOS 4 × 10−18 J 1 × 10−18 J

[30] UMTS 0.18µm CMOS 1 × 10−17 J 6 × 10−18 J
[30] GSM 0.18µm CMOS 2 × 10−17 J 6 × 10−19 J
[38] GSM 0.35µm CMOS 5 × 10−17 J 2 × 10−18 J
[32] Bluetooth 0.18µm CMOS 1 × 10−16 J 1 × 10−18 J
[2] AM 0.25µm CMOS 2 × 10−16 J 2 × 10−18 J
[2] FM 0.25µm CMOS 3 × 10−16 J 2 × 10−18 J
[7] speech 0.5µm CMOS 6 × 10−16 J 4 × 10−19 J
• the modification to the FOM reduces the variation in the achieved values because
the influence of VD D is eliminated: FOMDR,modif varies by a factor of 10 while for
FOMDR the ratio of the best to the worst FOM is about 100.
• the ranking especially differs for designs with a rather small input voltage: for
example [7] with only 56mV of input signal is ranked higher -thus fairer- using the
modified FOM.
Concluding the discussion on the power/performance analysis of the ADC the fol-
lowing use of FOMs is advised:
• for design purposes the elaborated FOMS of eq. 4.15 and eq. 4.16 can best be used
since they express a detailed relation between the various performance parameters
on the power consumption.
• for bench-mark purposes the conventional FOMs of eq. 4.1 and eq. 4.4 are easier
because they need less published data.
4.5 Power consumption in digital conditioning circuits

The functionality of the digital conditioning circuits is in:
1. decimating the incoming data, implying:
• the filtering of noise and interferers outside the wanted channel

Figure 4.3: Definition of the decimation filter characteristics
• the reduction of the high sample rate to the Nyquist sample rate.
2. reducing the dynamic range by scaling the word-length (digital VGA).
Both aspects are discussed next.
Decimation filtering: The decimation function is clarified based on fig. 4.3 depicting
the transfer function of the i th decimation filter (assuming multi-stage decimation). It
provides suppression of noise (shaped quantization errors of the ADC) and interferers
(remaining in case of weak analog filtering) that are present around the target sample
rate m i−1 f s . Next, the filtered signal can be re-sampled. The key specifications on the
filter are:
• the transition bandwidth f : in the i th decimation stage it equals f = m i−1 f s −

2BW. In the last decimation stage the transition band equals f = f s −2BW.
• the stop-band suppression δs : In a worst-case scenario all noise or interferer en-

ergy can be present in a single tone. Therefore, a suppression equaling the SINAD
specification is required.
• the pass-band ripple δ p
In the power/performance analysis only the stop-band suppression is included as a vari-

able. The pass-band ripple and the transition band in the last decimation stage are assumed
to be fixed by the application2 and do not take part in the balancing between analog and
2 In digital communications for instance, the Nyquist sample rate f is set by the symbol, bit or chip rate
s
(e.g. a 3.82M S/s chip rate for UMTS). The signal bandwidth BW is determined by the modulation index or by
the requirement of a guard band between neighboring channels. A different example is found in digital audio
systems. Here, the Nyquist sample rate is fixed for standardization reasons (e.g. f s = 44.1k S/s for digital
audio) while the signal bandwidth can be limited to the audible frequency range.
digital conditioning. In addition, the number of decimation stages and the decimation
factor are less important because these belong to the digital optimization. The analysis
assumes an optimal digital design and derives a power/performance relation that can be
used for extrapolation of the power consumption of this design to varying performance
requirements or vice versa.
Scaling: The second function is in reducing the dynamic range of the received signal
to the SINAD that is required for further processing, like for example digital demodulation.
In digital, this can be implemented in a rather straightforward way by simple bit-shifting
if a control signal (a Received Signal Strength Indication or RSSI) is available. There-
fore, it is not taken into account in the calculations. Instead, the decimation filtering is
emphasized.
4.5.1 Filter functions

Digital filters can be divided into two main classes: the finite impulse responses (FIR) and
the infinite impulse response filters (IIR). Although IIR filters can realize a sharp transi-
tion band at low complexity compared to FIR filters they are prone to coefficient growth
and stability problems. In addition, IIR filters suffer from phase distortion. This is a major
drawback obstructing the use of IIR filters in many applications.
Below, some often-used filter functions that are inherently linear-phase -or can be
designed as such- are briefly discussed. More detail can be found in [25], [66], [67], [68],
a.o.
CIC filters [69], [70]: Cascaded-Integrator-Comb filters consist of an integrator sec-

tion at the high sample rate and a comb section operating at the low sample rate. They
realize a frequency response:

sin(π fmT) k

H(f ) = with T = 1 (4.17)
m sin(π fT) mfs
k is the number of cascaded sections, m is the decimation factor. This frequency response
has zeros around multiples of the decimated sample rate. Hence, CIC filters efficiently
suppress the frequency bands that would fold back into the wanted channel when re-
sampling. They are often used in a first decimation step because they don’t need mul-
tiplying with coefficients but just a single addition per tap is made. Disadvantages are
in the restrictions on the pass-band width (it must be small compared to the decimated
sample rate to keep droop limited) and on the weak roll-off outside the pass-band. This
limits their application to decimation until four times the Nyquist sample rate.
Half-band filters [71]: Linear-phase half-band filters are characterized by a sym-

metrical transition band limiting their use to decimation by a factor of 2. Because of the
symmetry, they can be implemented with half the number of multiplications compared
to arbitrary filter designs. Hence, a half-band design is very attractive for use in the first
decimation stage running at the highest frequency.
Generic FIR filters: Especially in the final decimation stage a sharp transition band
and droop-correction need to be realized. Typically, this last stage needs the largest num-
ber of taps. On the other hand, it operates at the lowest frequency making the switching
dissipation acceptable. A generic FIR design can realize these requirements.

In general, the power consumption P of any digital block is determined by the supply
voltage VDD , the total capacitance C that needs to be switched, the clock frequency f and
the parameter α expressing the average activity of the gates:
P = αfCVDD
2
(4.18)
The aim of the analysis is in comparing power consumption of analog and digital signal
conditioning. Therefore, a relation between the power consumption in the decimation fil-
ter and the parameters ENOB, mfs and f s is mandatory3 . Contributions that do not scale
with these parameters can be disregarded. In addition, the activity α and the supply volt-
age VDD in eq. (4.18) are considered independent thereof4 . Hence, the prime challenge
is in expressing the capacitance C (dominated by the number of gates) in terms of the
quoted parameters. This is done in appendix C. In addition, the following upper limit on
the power/performance relation for digital decimation filters is derived there:
 
ENOB 3
Iworst−case ∼ O  f mfs ln m (4.19)
2fs
ENOB appears in a cubic relation: in a worst-case scenario, the number of filter taps, the
word-length of the data and the number of computations (in terms of partial additions)
are all proportional to ENOB. In the appendix, it is shown that for the first decimation
stages and for CIC implementations, a quadratic relation is more likely. Furthermore, the
highest sample rate mfs and the smallest transition band f /2 f s are assumed, while the
factor ln m is a measure for the number of consecutive decimation stages. Further on, it is
neglected because its influence is small.
4.5.3 Discussion
Again, the above results are further commented in the sense that the major influence of
the sample rate is highlighted. Also, the benefit of technology scaling is drawn to the
attention.
3 ENOB is used here instead of SINAD in the discussion on analog. For simplicity, it is assumed that all
bits are effective; i.e. we do not distinguish between the effective-number-of-bits (ENOB) and number-of-bits
(NOB).
4 The average activity is determined by the input data and the implementation of the filter. The supply voltage
is set by the technology. Sometimes, though, it is adapted according to the required processing speed. This
potential dependence on f s is disregarded in the present analysis. It can easily be taken into account afterwards.
Effect of sample rate

The derived results are valid if power consumption of the digital conditioning channel
is dominated by the decimation filters. In many digital designs, other contributions are
important as well. These are typically due to parasitics, e.g. wiring capacitance or due to
additional circuits, e.g. for test or control. In addition, the power consumed in the clock
distribution can be very important. Notice that this contribution scales with mfs .
In low power digital design it is common practice to adapt the supply voltage ac-
cording to the required sample rate5 . The variation of the supply voltage is limited by
the noise margin on the low side and by reliability issues on the high side. Within this
window, the supply voltage can be scaled to the sample rate. Using eq. 4.18, power con-
sumption then becomes cubic in mfs instead of proportional. In the above analysis, this
variation has been disregarded for simplicity reasons.
The various remarks in this paragraph do underline the prime importance of the sam-
ple rate on the power consumption of a digital block.
Technology scaling
Any analysis of power consumption in digital blocks would not be complete without
stressing the benefits of technology scaling. While for analog circuits technology scaling
hardly brings any power savings (in fact, beyond 0.25µm-CMOS a penalty in power
consumption is more likely [5]), for digital circuits the power decrease is substantial.
Porting the decimation filter from one technology generation to the next reduces the total
capacitance by a factor stech , with stech being the technology-scaling factor [72]:
• the area scales by 1/stech

2
• the gate-oxide thickness scales by 1/stech
Hence, the current consumption of the digital block scales accordingly:
Idigital −→ Idigital · stech (stech is the technology scaling factor) (4.20)
In addition, the supply voltage can be scaled by a factor ssupply . According to equa-
tion 4.18 this results in the following power saving:
Pdigital −→ Pdigital · stech · s2supply (ssupply is the voltage scaling factor) (4.21)
Assuming stech and ssupply both approximate 0.76 , per transition, some 65% of power
consumption is saved going from one technology generation to the next. From Moore’s
Law, it can then be derived that the dynamic power consumption of digital circuits is
reduced by a factor of 1000 every 10 years; i.e. in about 6 to 7 technology generations.
5 In fact, this technique can be actively applied to lower the power consumption of a digital design: using
parallelism in the implementation allows operation at a lower sample frequency and thus at a lower supply
voltage. The net effect is a decreased power consumption.
6 This has been true in the past but according to the ITRS roadmap, supply voltage scaling may be slowing
down in view of transition speeds.
4.6 Comparison
The power/performance relations derived for the major building blocks of the channel are
repeated here and discussed afterwards.
For analog conditioning circuits eq. 4.7 is valid:
BW SINAD1+α/4
I ∼ (α = 0, 1 or 2/3)
v̂IN2−α
Especially the α = 2/3 case is interesting: it is valid for circuits using local or global
feedback.
For the ADC eq. 4.14 is repeated:
BW SINAD7/6
I ∼ 4/3
v̂IN
For the decimation filter eq. 4.19 has been derived (neglecting the small influence of the
logarithmic function):
 
ENOB 3
Iworst−case ∼ O  f mfs ln m
2fs
Notice, the results are quoted in terms of SINAD or ENOB (instead of IM3 and DR) to
allow easy comparison between analog and digital.
Comparison of the power/performance relations

The power/performance relation of the ADC resembles that of analog conditioning
circuits. Below, some minor differences are explained. Next, these power/performance
relations are compared to the power/performance relation of digital conditioning circuits,
leading to some generic thoughts on the digitization of the conditioning channel.
Power/performance relation for A/D conversion versus that for analog condi-
tioning: The power/performance relation of the ADC is the same as that of the analog
conditioning circuits with α = 2/3. Therefore, the FOMs of eq. 4.15 through eq. 4.16 that
were derived for a ADC, are applicable to analog circuits as well. This results from
the assumption that the ADC power consumption is dominated by its analog circuits and
that these consist of similar differential pairs as those used for the analog conditioning
circuits. However, the proportionality factor between the left and right-hand side of the
above relations differs for two reasons:
• Different linearity: The ADC feeds back a pulse-density modulated signal

instead of a simple analog signal. Therefore, distortion in the input stage of the
ADC is twice larger than that of the analog circuits (eq. B.8 versus eq. 4.12).
This is detailed mathematically in appendix D.
4.6. COMPARISON 53
• Contribution of cascaded stages: The ADC as well as the analog conditioning

part consists of a cascade of sections all adding to the overall current consumption.
In both cases, the consecutive stages can have increasing noise and distortion be-
cause of the preceding gain. This is especially true for the ADC where the
loop filter can have very high gain at low-frequencies. Because of the closed-loop
operation, the filter only needs to meet a gain/phase criterion at high frequencies.
On the contrary, the design of the analog cascade is often much more constrained.
For instance, the transfer function is tightly specified and, in the likely case of
open-loop operation, the gain of the consecutive stages is limited. Because of these
constraints, a cascade of several analog sections is likely to consume more than the
loop filter of a ADC of the same order.
The latter observation is supported by the example of an FM receiver in chapter 8. Ta-

ble 8.5 compares the power consumption of a highly analog conditioning channel (first
column) to that of a the ADC in a highly digital channel (last column).
Power/performance relation of digital conditioning versus that for analog condition-

ing or A/D conversion: Power consumption in the decimation filter is proportional
to ENOB or a power of ENOB while in analog circuits and ADCs it is an exponen-
tial function of ENOB (i.e. linear in SINAD). Even under the worst-case assumption of
eq. 4.19 the power increase as a function of ENOB is moderate compared to that in analog
circuits.
Example: conditioning channel for FM receiver (section 8.3)

In fig. 4.4 the power/performance relations are applied to an example conditioning chan-
nel. In the referenced implementation, the conditioning channel achieves 14 bits of reso-
lution and the power consumption of the ADC and that of the digital decimation filter
are about equal. From this measurement point, the power consumption is extrapolated
as a function of ENOB. In this example, Panalog represents the power consumption of the
ADC. Pdigital is extrapolated both as a square and as a cubic function of ENOB in
order to demonstrate the minor difference.
In general, equations 4.9, C.6 and C.7 can only be used for extrapolation purposes be-
cause they express a proportionality. Then, the crossing of the power/performance curves
depends on the actual implementation and on the various parameters in eq. 4.9, C.6 and
C.7.
Conclusions with respect to digitization: The power consumption of the complete

channel equals the sum of the analog (or ADC) power and the digital power consumption
and, in fig. 4.4, is represented by the dotted envelope curve. From this example, it can be
generalized that:
• in channels targeting a low ENOB-value, the power consumption of digital blocks

is likely to be dominant
Figure 4.4: Power consumption of analog and digital blocks as a function of

ENOB (normalized at 14 bits corresponding to an example in sec-
tion 8.3)
• in channels targeting a high ENOB-value, the power consumption is dominated by

the analog blocks
In a generalized version of fig. 4.4, Panalog would either represent the power consumption
of the analog circuits (in case of a highly analog channel) or it is the power consumed in
the ADC (in case of a highly digitized channel). It is not self-evident that digitization
of the channel implies a power reduction. On one hand, this may be the case because:
• while the same power/performance relation is valid for analog circuits and the
ADC, the absolute power consumption of a high-order ADC, likely, is
lower than that of a cascade of several analog circuits (see discussion on page 53
on cascaded stages)
On the other hand, though, digitization may just as well lead to a huge power increase for
the ADC:
• due to the high resolution and bandwidth demands on the ADC, the limitations of
section 3.4 are hit and its power consumption increases faster than what is predicted
by eq. 4.14
In addition, the power consumption of the digital part increases. The power consequences
of digitization are further explored in chapters 5 and 6 for various conditioning channels.
In general, digitization of the conditioning channel does not necessarily imply a power
saving.
Finally, considering:
• the ever-increasing resolution demand, i.e. a higher ENOB target

4.7. CONCLUSIONS 55
• the benefits of technology scaling, i.e. the Pdigital -curve shifts down
it can be concluded that:
• power consumption of analog blocks becomes dominant in more and more condi-
tioning channels
• this is also true in highly digitized channels: there, the “analog” power is consumed
in the ADC
These observations underline the growing, paramount importance of low-power analog
circuit design. However, the power/performance relation of analog circuits only im-
proves gradually in time. In order to safeguard an acceptable power consumption for
high-resolution conditioning channels in advanced technologies, architectural innovation
is needed. Current research approaches on the architectural level include:
• digitally tuning/assisting analog circuits [20]
• dynamic scaling of power consumption to the instantaneous performance target
• reducing the number of analog stages (i.e. the number of current branches) by merg-
ing functionality
Especially the latter two techniques are illustrated in chapter 6, introducing conditioning
ADCs, and in the various designs presented throughout chapters 7 and 8.
Comparison to literature
The curves of fig. 4.4 correspond to the linear and logarithmic relation as a function of
SINAD derived by a.o. Vittoz [55] and Meindl [56] for power consumption in analog and
digital respectively. However:
• the referenced papers concentrate on fundamental limits for power consumption of
a single analog or digital cell
while here:
• far more elaborated relations for the estimated consumption of actual channels are
derived;
• in addition, the derived relations are used further on to improve the overall power/per-
formance ratio of the channel crossing the traditional boundaries between analog,
mixed-signal and digital design.
4.7 Conclusions
Because of the lack of analytical power/performance relations for analog and digital con-
ditioning blocks in literature, a dedicated analysis is performed in this chapter. This analy-
sis results in the observations listed next.
The power consumption of a single-bit continuous-time ADC can be extrapo-

lated to a modified performance target, using the following Figure-of-Merit:
4/3
P v̂IN
FOM =
SINAD7/6 BW VD D
On the other hand, for ranking purposes, more common FOMs can be used because they
require less parameters.
It is shown that the power/performance relation of analog circuits is similar to that
of a continuous-time ADC. The absolute value of the power consumption of a high-
order ADC is likely to be lower than that of a cascade of several analog circuits
because of the overall feedback.
Consequently, digitization of the channel -where the analog circuits are replaced by a
difficult ADC and consecutive digital conditioning- does not self-evidently yield a power
saving, nor does it necessarily yield a power penalty.
The power consumption of a decimation filter is only a weak function of the ENOB
specification (when compared to the relation for an analog function) but it is strongly
affected by the sample rate.
The power consumption of analog circuits or of the A/D converter becomes dominant
in most conditioning channels. This is also true in highly digitized channels.
A power reduction in the conditioning channel requires architectural innovation of
analog blocks or of the ADC.
These observations are used further on for exchanging analog and digital conditioning in
a channel.
Chapter 5
Full-analog and full-digital

conditioning channels
Building on the results of the previous chapter, two channel architectures are compared in
terms of their power/performance ratio. The channels differ in the degree of digitization
of the signal conditioning1 . The first architecture is a conventional channel with full-
analog signal conditioning. The second channel architecture has full-digital signal con-
ditioning and therefore is highly flexible. The full-digital conditioning channel requires
multi-channel A/D conversion at a very high sample rate. It is shown that, depending on
the application, a full-digital conditioning channel may not be feasible in view of current
ADC performance and with the present technologies. Especially, the requirement for
a very high sample rate is demanding, both for the ADC and for the digital part.
Hence, alternative channel topologies are required in order to maintain a competitive
power/performance ratio while pursuing digitization. This is the topic of chapter 6.
5.1 Full-analog conditioning channel

The operation of the channels (in this and the next chapter) is clarified with the exam-
ple input signal of fig. 5.1.a. It consists of a small wanted channel, a slightly stronger
adjacent channel and two much stronger far-off interferer channels that could cause inter-
modulation components in the wanted channel. At the output of each conditioning chan-
nel, only the digitized wanted channel is available with a specified SINAD and sample
rate (fig. 5.1.b). As a starting point, a channel with full-analog conditioning is discussed.
Further on it serves as a benchmark for the power/performance ratio of digitized channels.
1 The term “signal conditioning” is used here in the limited sense; i.e. it only refers to analog/digital channel
filtering and analog VGA or digital word-length scaling (see nomenclature defined in fig 1.3)
57
58 FULL-ANALOG AND FULL-DIGITAL CONDITIONING
Figure 5.1: Example input spectrum (a) and output spectrum (b) of the condi-
tioning channel
5.1.1 The conditioning channel

Fig. 5.2 depicts a block diagram of a conventional, highly-analog conditioning channel
with A/D conversion. The bottom graph of fig. 5.2 shows the conditioning of the
signal in the various stages. The parameters and symbols that are used here have been
defined in section 2.5. DRi indicates the value of the input-referred dynamic range for
the wanted channel. SINADo is the value of the SINAD in the bandwidth of the wanted
channel available at the output of the conditioning channel. If the subscript is omitted,
it is explicitly specified to which bandwidth and position in the channel DR and SINAD
refer.
• the channel filter Hch (s): reduces the signal bandwidth by a factor p resulting in
a p-times lower integrated noise power and suppresses the large interferers with
overall2 rms-value q vMAX
• the VGA: reduces the amplitude range of the input signal by varying the gain be-
tween a maximum (GMAX ) and a minimum value (GMIN ). This is typically allowed
in a receiver-type of application. For simplicity, the output-referred noise is as-
sumed to be the same in all gain settings in the remainder.
• the ADC: digitizes the signal adding a lot of quantization noise and increasing
the bandwidth requirements due to the over-sampling.
• the decimation filter Hdec [z]: suppresses the shaped noise and reduces the sam-
ple rate such that at its output the specified SINADo and signal bandwidth become
available.
Clearly, the filter Hch (s) is the most challenging block in this architecture: it must achieve
low noise and distortion in the presence of a wide-band input signal with a large ampli-
tude range. The requirements on the subsequent blocks become more and more relaxed.
2 It is assumed that the overall amplitude of the sum of all interferers is a factor q larger than that of the
maximum wanted channel. If this is not the case then q should be put to 1.
5.2. FULL-DIGITAL CONDITIONING CHANNEL 59
Figure 5.2: A conventional conditioning channel with ADC
Not visible in this diagram are the bandwidth requirements on the analog circuits
throughout the channel. The bandwidth of the filter circuits must often be large compared
to the entire signal bandwidth (including interferers). The bandwidth of the analog cir-
cuits in the ADC is determined by the sample frequency m 1 f s .
An example implementation of a conventional conditioning channel for a FM re-
ceiver is presented in chapter 8.
5.2 Full-digital conditioning channel

As a counterpart to the full-analog channel, multi-channel A/D conversion and full-digital
signal conditioning are analyzed next.

The full-digital conditioning channel of fig. 5.3 offers great flexibility (to accommodate
changing standards or multi-channel and multi-mode operation) and allows easy technol-
ogy scaling (to shrink the area or power consumption). The burden of this solution is
on the ADC. Both the interferers -in adjacent or in far-off channels- and the wanted
channel are digitized (fig. 5.4). Because of its inherent alias-filtering, the continuous-time
ADC is robust to high-frequency products at its input. For example, these may have
been generated by a mixer that precedes the channel of fig. 5.3. Table 5.1 compares the
requirements on the multi-channel ADC to those on the ADC in the full-analog
channel. The multi-channel ADC needs to convert a p-times larger bandwidth. Its stable
input range must be dimensioned to the amplitude of the dominant interferer, even while
the wanted channel may be small due to the lack of preceding gain. As such, both a high
Figure 5.3: Full-digital conditioning channel with multi-channel A/D con-

version
Table 5.1: Specifications on the ADC in the full-analog and in the full-
digital channel
Specification ADC in ADC in

full-analog channel full-digital channel
SINAD SINADo q 2 DRi / p

bandwidth BW p BW
maximum input G MIN vMAX q vMAX
√ √
intermodulationa 2/ SINADo 2/q DRi
a referred to the maximum input signal of the ADC, using eq. 2.2
IM3 is necessary because of the large interferers and the noise must remain low compared
to the smallest wanted signal. Note that these strict requirements are not caused by the
wanted channel but are due to the strong interferers. The above requirements result in a
more than p-times larger sample rate: next to the higher conversion bandwidth, it is very
likely that a higher over-sample ratio and loop filter order are mandatory for the ADC to
meet the signal-to-quantization-noise ratio. Possibly, the specifications on the ADC
rise above what is feasible in view of state-of-the-art performance (see limitations in sec-
tion 4.4).
Figure 5.4: Output spectrum of the multi-channel ADC in the full-digital

channel
Finally, the channel selection is performed in the digital domain in the filter Hch [z]
and the word length can be scaled in the block G[z] (i.e. digital VGA). In practice, Hch [z]
and G[z] are merged with the decimation filter Hdec [z].
5.2.2 Power/performance analysis

First, the power consumption of the multi-channel ADC is compared to that of the
ADC in the full-analog channel. Next, the overall power consumption of the full-
digital channel is analyzed relative to that of the full-analog channel (fig. 5.2). For com-
pleteness, the requirements on peripheral circuits (e.g. reference circuits) in both channels
are briefly compared as well.
The ADC
The multi-channel ADC is compared to the simple ADC in the full-analog chan-
nel. The analysis is conducted from two angles being resolution (i.e. noise and distortion)
and bandwidth requirements.
Noise and distortion: The power/performance relation of eq. 4.13 and the require-
ments specified in table 5.1 are used to compare the current consumption IADC (digital CCh)
of the ADC in the full-digital channel to the current IADC (analog CCh) of the ADC
in the full-analog channel. The abbreviation CCh indicates “conditioning channel”:
7/6
IADC (digital CCh) DRi
=q G 4/3
MIN
(5.1)
IADC (analog CCh) SINADo
Figure 5.5: Comparison of a full-analog and a full-digital conditioning chan-

nel in terms of power/performance
The lack of analog filters or VGA results in a very large resolution requirement on the
ADC and a consequent power increase. Especially in wireless communication sys-
tems, the difference between DRi and SINADo can be very large. A further increase is due
to the q-times larger overall amplitude because of the interferers while the wanted signal
is G MIN times smaller.
Bandwidth: The multi-channel ADC requires a much larger circuit bandwidth:

• the bandwidth of interest is p times larger
• the over-sampling factor m 2 must be higher in order to achieve the required SQNR
From both angles, the multi-channel ADC is significantly more difficult -and there-
fore more power-hungry- than the ADC in the conventional channel.
The conditioning channel

The above power increase for the ADC should be put into perspective (see sec-
tion 4.6). In fact, the power consumption of the overall channel should be compared
to that of the conventional channel instead of only comparing the ADCs. This analysis
is split into two steps (see fig. 5.5). First, the power consumption of the multi-channel
ADC is compared to the joint power consumption of the analog part of the conven-
tional channel. Next, the digital part of both channels is compared.
Analog part: The multi-channel ADC is compared to the cascade of analog

circuits and ADC in the conventional channel. On one hand, noise and distortion
specifications on the multi-channel ADC are the same as those on the input stage -i.e.
the filter- of the conventional channel. On the other hand, the power/performance depen-
dence of the multi-channel ADC is similar to that of the input stage of the conven-
tional channel (assuming the input transconductor dominates performance, see chapter 4).
Therefore, one can argue that, power-wise, the ADC in the full-digital architecture is
competitive to the input stage of the conventional channel.
Moreover, in the discussion in section 4.6 it was argued that the consequent stages
of the ADC are likely to consume less than the corresponding analog stages in the
conventional channel. In addition, the discussion in section 3.3.2 demonstrated that the
accuracy requirements on the analog sections in the conventional channel are strict as
compared to those on the consequent stages of the ADC. Hence, it may seem that the
power consumption of the multi-channel ADC could even be lower than that of the
analog part of the conventional channel. However, up to now only noise and distortion
have been considered. The true challenge on multi-channel A/D conversion lies in sample
rate related problems.
As discussed above, the sample rate of the multi-channel ADC may become very
high. The bandwidth of the sub-circuits of the multi-channel ADC must be related to
this high sample rate while the bandwidth of the analog circuits in the conventional chan-
nel is only related to the -much smaller- signal bandwidth. Hence, for bandwidth reasons
the current consumption of the multi-channel ADC is likely to be much higher than
that of the input stage of the conventional channel.
Digital part: With m 1 being the over-sampling factor for the conventional and
m 2 the over-sampling factor for the multi-channel architecture, the following worst-case
power increase is to be expected (eq. 4.19 but neglecting the small influence of the loga-
rithmic function):
3
Idigital (digital CCh) m 2 p log DRi q 2 / p
= 3
(5.2)
Idigital (analog CCh) m 1 log SINADo
The higher resolution demand only puts a relatively small penalty on the power consump-
tion in digital because of the logarithmic relation. On the contrary, the increase in clock
rate has a major effect. So, also from a digital perspective multi-channel conversion is not
necessarily an attractive option yet. Of course, scaling laws are beneficial here.
The discussion is illustrated with an example. A Bluetooth application is considered

because -when compared to UMTS, GSM, a.o.- this is a relatively “easy” application.
Therefore, it may be a candidate for digitization.
Example: Bluetooth
A Bluetooth channel is about 1MHz wide, the entire band spans 78MHz (hence p = 78).
Digital demodulation requires a resolution of about 18dB. The amplitude range of one
channel is 50dB. The maximum interferer amplitude is as large as the maximum wanted
signal, i.e. q = 1.
Taking sufficient margin, we compare:
• an ADC for the full-analog conditioning channel, with SINAD = 40dB (i.e. 18dB)3
realized using m 1 = 16 and L = 3
3 Comparison to an ADC with only 18dB of SINAD is unrealistic. First of all, because that requires a lot
of variable gain and gain control and therefore becomes un-practical. Secondly, it is unlikely that a ADC
would be used for such low requirements.
• an ADC for the full-digital conditioning channel, with SINAD = 70dB (i.e. >
(18 + 50)dB) realized using m2 = 40 and L = 4 to 5
Then, for noise and distortion reasons, eq. 5.1 predicts:

7 7/6
IADC (digital CCh) 10
=1 104/3 ∼
= 1e5
IADC (conventional) 104
Still, it is expected that the calculated power increase in the ADC corresponds to what is
gained by omitting the analog conditioning circuits. Instead, the dominant power penalty
is due to higher bandwidth requirement. In the analog channel the ADC is sampled at
32MHz, in the digital channel at 6.24GHz. This is far beyond what is feasible at present
(see table 3.2).
The current consumption of the decimation filter increases according to eq. 5.2:
3
Idigital (digital CCh) 40 · 78 log 107 /78 ∼
= 3 = 400
Idigital (analog CCh) 16 log 104
The input sample rate of 6.24Gs/s is hardly feasible in present-day CMOS technology and
at least not common for DSPs.
This example shows that, even for a standard as “easy” as Bluetooth, a full-digital
signal conditioning channel does not seem feasible in view of state-of-the-art ADC
performance and speed of DSP.
Analog reference circuits

Although the power consumption in the peripheral circuits (such as time, voltage or cur-
rent references) is not systematically analyzed in this book it is worth putting a remark
here. These circuits generate a reference for the ADC (i.e. for the feedback DAC).
Their accuracy must be as good as what is required of the ADC as a whole. Clearly,
the multi-channel ADC requires high-precision and, by consequence, power-hungry
references. Especially, the clock-generation is extremely challenging: not only does it
need to have low jitter, it must also provide a high frequency.
From the discussion and the example, it becomes clear that full-digital signal con-
ditioning puts impressive requirements on the ADC. In view of state-of-the-art ADC
performance and DSP capability, full-digital signal conditioning is not yet feasible for
systems operating over a large frequency band with a large dynamic range. Hence, this is
the case for most wireless systems.
An exception is presented in chapter 8. There an AM-radio receiver with multi-
channel A/D conversion is discussed. AM radio is a narrow-band system. As such, the
sample rate of the multi-channel ADC does not become excessive. Instead, it enables a
power-effective alternative to a full-analog channel and provides a lot of flexibility.
5.3. CONCLUSIONS 65
5.3 Conclusions
Full-digital signal conditioning with multi-channel A/D conversion puts impressive re-
quirements on the ADC. The sample rate of the ADC and the digital circuits increases
dramatically, possibly exceeding state-of-the-art sample rates for a ADC and com-
mon DSP speed.
Straightforward digitization results in a major increase in the power consumption of
the ADC and of the overall channel.
Full-digital signal conditioning is not yet feasible for systems operating over a large
frequency band with a large dynamic range. This is the case for most wireless systems.
The discussion in this chapter is illustrated with example implementations in chapter 8.

Chapter 6
Conditioning ADCs
Because of the over-sampling and the continuous-time loop filter, a certain level of inter-
ferers can be applied to the ADC without corrupting the resolution in the bandwidth
of the wanted channel. This characteristic of the continuous-time ADC is exploited in
a third architecture for the conditioning channel. The ADC is used without preced-
ing analog filters or VGA. Instead, its inherent interferer immunity is exploited or even
enhanced. This is achieved by moving some signal conditioning into the loop. The
ADC, used in this manner, is referred to as a “conditioning ADC”.
First, the concept of a “conditioning ADC” is introduced along with an overview
of architectures. Next, a universal model of a generic ADC is described. This model
is used to analyze the major limitations on the allowable interferer level for the ADC.
Then, various topologies for the conditioning ADC are presented. These topologies
differ in the amount of signal conditioning that is integrated into the loop and in the
power/performance relation of the overall channel. They are compared along the lines
developed in chapter 5.
6.1 Generic conditioning ADC

In chapter 5 channels with full-analog and full-digital signal conditioning have been pre-
sented (fig. 6.1.a and b respectively). Here, the signal conditioning is integrated into the
ADC resulting in a “conditioning ADC” without preceding analog filters or vari-
able gain stages (fig. 6.1.c). First, the basic operation of the conditioning ADC is
clarified. Next, a universal model of a ADC is introduced. This model is used for a
generic analysis of the interferer immunity of continuous-time ADCs. At the end of
this section, a basic power/performance analysis is presented.
6.1.1 Concept of operation

The concept of implementing some signal conditioning into the ADC, is developed in
three steps. The solutions presented in sections 6.2 and 6.3 exploit characteristics of
67
68 CONDITIONING ADCs
Figure 6.1: Full-analog (a), full-digital (b) and mixed-signal (c) signal condi-
tioning
known topologies. The solution presented in section 6.4 realizes the signal-conditio-
ning by explicit design. The consecutive solutions build on each other’s properties. They
are introduced only briefly at present but are detailed later in this chapter.
Signal conditioning in the decimation filter

Like in the full-digital architecture of section 5.2 both the wanted and the interferer chan-
nels are applied to the ADC, i.e. no analog pre-filtering or VGA is applied. A key
difference compared to the full-digital architecture is the fact that the conversion band-
width of the ADC is only as large as the bandwidth of the wanted signal. The operation
of the conditioning ADC is intuitively illustrated in fig. 6.2 for the same composite
input signal as used in chapter 5. The basic characteristic enabling this operation, is the
fact that a ADC uses over-sampling. On top of that, in case of a continuous-time
implementation, the loop filter acts as an anti-alias filter. This results in some immunity
of the ADC to interferers; i.e. interferers -below a set amplitude limit- can be applied
without affecting the resolution in the bandwidth of the wanted channel. The limit on the
immunity is analyzed in a more quantitative way in section 6.1.3.
Using the ADC in this way, the interferer channels appear in the noise-shaped
part of the output spectrum. The ADC provides frequency selective resolution; i.e. the
6.1. GENERIC CONDITIONING ADC 69
Figure 6.2: Input spectrum (a) and output spectrum (b) of a conditioning
ADC
wanted channel is digitized accurately while interferer channels are de-correlated by the
shaped quantization noise. Next, the decimation filter attenuates the interferers, along
with the quantization noise. This is illustrated in fig. 6.3.a: the interferers pass through
the ADC1 but are attenuated by Hdec [z]. The latter also includes the digital VGA.
Note that, while in the full-digital architecture of fig. 5.3 a dedicated digital channel
filter Hch [z] is needed, here, this functionality is inherent to the decimation filer Hdec [z].
At the input of the decimation filter, the “unwanted energy” now includes both the quan-
tization noise and the interferers. Still, the total “unwanted energy” is the same as in the
situation without interferers. It equals the energy in the bitstream, which is a constant, mi-
nus that in the wanted signal. Thus, it is independent of the interferers. On the contrary,
the distribution of the “unwanted energy” can be different. Still, in practice, this should
hardly affect the attenuation requirements because the decimation filter must anyway be
designed to suppress worst-case tones. Similarly, the word-length scaling that happens in
a decimation filter, simultaneously fulfills the role of digital VGA.
Finally, notice the following:
• while the signal-conditioning happens in the decimation filter, still, this solution is
classified as a conditioning ADC. Since the decimation filter is indispensable
for A/D conversion, it can be considered part of the “broader ADC”;
• the signal-conditioning in the decimation filter is only possible because of the in-
terferer immunity of the continuous-time ADC (see section 6.1.3).
In section 6.2, the channel with signal-conditioning in the decimation filter, is analyzed
for the example of a feed forward ADC, as a reference2 .
While having economized on the channel filter and VGA, some analog circuits in
1 Note that, in this example, the interferers are amplified throughout the ADC; i.e. the transfer of the
interferers of amplitude q v̂MAX shows a weak, positive slope. This is due to the frequency dependence of the
signal transfer function (STF) of the ADC. The inclination and direction of the slope should be considered
as an example since they depend on the STF of the ADC and on the frequency of the interferers.
2 In fact, any continuous-time ADC can be used in this way because this conditioning channel is enabled
by the over-sampling that is inherent to modulation and by the anti-alias suppression of the loop filter
Figure 6.3: Conditioning ADCs with signal-conditioning in the decimation

filter (section 6.2) (a) or using a filtering STF (sections 6.3 and 6.4)
(b)
this channel architecture remain very difficult. This is alleviated in the following solu-
tions using signal-conditioning in the ADC itself (next to conditioning in the decimation
filter).
Signal conditioning with restricted filtering STF
Next to its inherent interferer immunity, a second characteristic of a ADC can be

exploited; i.e. the signal transfer function (STF) of a ADC is not flat, instead, it shows
frequency selectivity. If the frequency selectivity is favorable -i.e. |STF| < 1 for interferers
channels while |STF| = 1 for wanted channels- this property can be used for implementing
signal-conditioning into the ADC. In section 6.3, it is shown that ADCs based on a
loop filter with nested feedback, fulfill this requirement. Then, the STF provides filtering3
from the input to the output of the ADC. In addition, VGA functionality is easily added
either by varying the input resistance or the full-scale level of the feedback DAC. Note
that the achieved filtering is restricted in the sense that the STF and the noise transfer
function (NTF) cannot be optimized independently.
The signal-conditioning is further illustrated in fig. 6.3.b. The interferers at an input
level q v̂MAX are attenuated throughout the ADC. On the contrary, the wanted signal
is amplified with a gain ranging from G MIN to G MAX depending on the signal strength.
Hence, filtering and VGA are integrated into the ADC and the same functionality is
realized as in the conventional cascade of filter, VGA and ADC in fig. 5.2. It can be
concluded that, if the STF of the ADC is favorable, a true “conditioning ADC”
can be implemented.
Signal conditioning by unrestricted STF design
While, in the above solution a given STF of a known topology, is exploited for signal-
conditioning, the STF of the ADC may as well be designed for signal-conditioning. Ded-
icated design of the STF, allows for optimization according to the desired filtering char-
acteristic. Contrary to the above solution with restricted filtering, this is possible without
compromising on the noise shaping. In addition, a more power-efficient topology can
then be used. A conditioning ADC with this property is presented and analyzed in
section 6.4. It is referred to as a “filtering-feedback ADC” or FFB-ADC. The diagram
of fig. 6.3.b. remains valid for this ADC.
All solutions with a conditioning ADC rely on the inherent interferer immu-
nity of a continuous-time implementation. In fact, it is shown in section 6.3 and
6.4 that the solutions with signal-conditioning integrated into the ADC are enabled by an
improved interferer immunity of the design.
3 “Filtering” is used as a synonym for “frequency selectivity”. In this book, this term does not imply full
attenuation of interferers.
Figure 6.4: Universal model of a 1-bit modulator (a) and linearized equiv-
alent (b)
6.1.2 Universal model of a modulator

A universal model of a single-bit modulator is depicted in fig. 6.4.a [73], [25]. This
model captures all topologies that are discussed further on and therefore enables a
generalized discussion4 . It consists of a linear block comprising the filter functions L 0
and L 1 and a non-linear block consisting of the 1-bit quantizer. The DAC in the feedback
path is implicit. L 0 and L 1 are filters acting on the input signal and on the feedback
signal respectively. A simple linearized model is shown in fig. 6.4.b where the quantizer
is replaced by a linearized gain c and an additive noise source E n . A calculation of c and
E n is a.o. documented in [25], section 4.2.1. The signal transfer function of the ADC
can be calculated in the s-domain:
Y (s) c L 0 (s)
= (6.1)
X (s) 1−c L 1 (s)
The noise transfer function (NTF) equals:
Y (s) 1
= (6.2)
E n (s) 1−c L 1 (s)
Numerous publications treat the optimization of L 1 to improve the NTF while maintaining
stability [74], [75], a.o. On the contrary, relatively little attention has been paid to the
STF of the modulator. Publications that do treat STF-design especially target a flat
pass-band response for audio applications [76], [77]. Here, the STF-design is exploited
to improve the immunity of ADCs to interferers outside the ADC pass-band.
The model of fig. 6.4 is introduced in view of a generic analysis of various top-
ologies introduced further on. All those topologies realize the same L 1 and thus the same
NTF (given a fixed order for the loop filter). The loop filter L 0 is designed to optimize the
STF in view of interferer immunity.
6.1.3 Interferer immunity

The over-sampling principle used in ADCs is the basic reason for their robustness
to interferers outside of the conversion bandwidth. As the amplitude of the interferers
4 Here, this model is meant for describing only the ADC. More generalized, it can be used to include the
analog conditioning in front of the ADC as well. Then, this functionality would be part of the L 0 (s)-function.
Figure 6.5: ADC (a) and linearized model (b) for analyzing the anti-
aliasing suppression
exceeds a set limit, though, the interferers do cause distortion, spurious responses and an
increase of noise in the wanted channel. This limit is due to various effects and can be
frequency dependent. These effects are analyzed and quantified next.
Aliasing
Compared to ADCs sampled at the Nyquist rate any ADC benefits from reduced
anti-aliasing requirements because of the over-sampling:
• the number of frequency bands that fold back into the conversion band is less than
in case of sampling at the Nyquist frequency.
Only continuous-time implementations benefit from an additional characteristic:
• the interferers near the clock frequency are suppressed by the loop filter before
being sampled in the quantizer.
The latter is not valid for discrete-time implementations: by definition, the transition to the
discrete-time domain -i.e. the sampling action- occurs already at the input of the discrete-
time ADC. The aliasing that occurs at this point is irreversible: the in-band spurious
cannot be removed by the loop filter.
A quantitative analysis of the alias-suppression is conducted based on fig. 6.5.
Suppose an interferer at a frequency mfs − f near the sample frequency is applied to
the continuous-time ADC. Typically, the unity-gain frequency of L 1 is at 1/6th of
the sample rate [25] for stability reasons. This implies that the feedback is not effective
anymore at the interferer frequency and the feedback path in fig. 6.5.a can be neglected.
The interferer is attenuated by loop filter L 0 and consecutively sampled in the quantizer.
This introduces an alias Z at f . Fortunately, this alias Z is suppressed towards the
output because of the high preceding gain at frequency f . The (approximate) transfer
of the alias Z ( f ) towards the output Y ( f ) of the ADC can be calculated from the
simple linearized model of fig. 6.5.b. The overall suppression then equals:
Y ( f ) c L 0 (mfs − f )
| |=| | (6.3)
X (mfs − f ) 1−c L 1 ( f )
L 0 (mfs − f )
≈| | (6.4)
L 1 ( f )
As mentioned, L 1 is kept constant in the remainder while various L 0 -functions are con-
sidered for improving a.o. the alias-suppression.
Stable input range
In section 3.2.3 the stable input level was defined implicitly, assuming only a signal within
the conversion bandwidth of the ADC is applied. Here, this definition is extended to
inputs outside the conversion bandwidth. For these frequencies, the stable input range
is defined as the input level for which the maximum modulation depth of the output is
reached. As such, this definition is an extrapolation of that in section 3.2.3, where it was
noted that, when applying an input amplitude as large as the stable input level, then the
modulation depth of the output is maximal.
As motivated in section 3.2.3 the maximum modulation depth of a well-designed
single-bit modulator is about 70%. This corresponds to a value of −3 dB compared
to digital full-scale. In the analog domain and assuming a DAC gain of 1, this is a value
of 0.7 v D AC . v D AC indicates the reference voltage of the DAC. The corresponding input
level, i.e. the stable input range, can then be derived by means of the STF of eq. 6.1:
0.7 vDAC
Stable input range(jω) = (6.5)
|STF(jω)|
Although the maximum output modulation depth is constant over frequency5 , the stable
input range is not. Its frequency dependence is inversely proportional to that of the STF.
Therefore, it can be improved by proper design of L 0 . (As mentioned, L 1 is kept constant
to realize optimal noise shaping.)
Notice that the above limit is derived from a linearized model of the ADC. As
such, the results are approximate and only valid within the small-signal operation regime.
Verification of this limit from the linear analysis with measurements in chapter 9 yields
good correspondence. Hence, the model is adequate for determining the transition to
instability and large-signal behavior.
5 In fact, for higher signal frequencies the maximum modulation depth of a bit-stream code increases slightly.
This effect is marginal and is disregarded in the remainder.
Spurious responses
The phenomenon of “spurious responses” in a modulator is due to the fact that its
quantization noise is not random. Instead, patterns -correlated to the input signal- may
appear at the output of the ADC in some, often not obvious way. In the frequency domain,
this translates into tones.
In this book, the term “spurious responses” is used in a limited sense. Here, we fo-
cus on the tones that are generated when an interferer is applied near even sub-harmonics
of the sample frequency, e.g. mfs /2, mfs /4, etc. In practice, the analysis can be further
narrowed to inputs applied near mfs /2 since these are most notorious.
First, available literature on the analysis of spurious tones is referenced. Next, the
mechanism causing spurious responses in a ADC is briefly recapitulated and some
techniques to reduce spurious tones are listed. Finally, it is noticed that, in practice, many
spurious responses may just as well be due to deficiencies in the analog circuits or the
layout rather than being caused by the tonal behavior of encoding.
Analysis: Spurious responses are well-studied for low-order modulators and DC-
inputs [78], [79]. For higher-order modulators the results become approximate and, again,
assume DC-inputs [80]. To the author’s knowledge, no analytical studies on spurious re-
sponses of high-order modulators due to high-frequency inputs have been published yet.
The discussion in this section remains qualitative as well. It focuses on pragmatic guide-
lines to reduce spurious responses caused by high-frequency interferer inputs.
Spurious responses for low-frequency inputs: This phenomenon is intuitively under-

stood from the theory on asynchronous modulators in [81] and considering the effect
of sampling next.
First, when applying a low-frequency input to an asynchronous modulator, the
following occurs:
• apart from the wanted output, also odd harmonics are generated;
• the input causes FM-modulation of the idling frequency of the modulator. This
results in a Bessel-spectrum around mfs /2 and around harmonics of this frequency.
Second, considering the sampling in a synchronous ADC [80], [82]:
• aliasing of the spurious tones occurs.
For this example of a low-frequency input, the latter means that a.o. the Bessel compo-
nents around the second harmonic of the idling frequency appear at baseband. Fortunately,
all the described tone-generating mechanisms occur in the quantizer. By consequence, the
tones appearing at baseband are suppressed by the noise shaping of the loop.
Notice that, for a small input signal, the amplitude of the spurious tones varies
monotonically with the input amplitude. This is obvious for the odd harmonics but is
also true for the Bessel-components. In fact, under this condition, especially the 1st -order
Bessel component is generated around mfs /2. Moreover, the amplitude of this component
is more or less linear with the input amplitude. For larger inputs, more and more Bessel
components become important. In addition, their amplitude varies very irregularly as a

function of the input amplitude.
Spurious responses for high-frequency inputs: In the present context of interferer

immunity of a ADC, we focus on inputs applied near mfs /2. This input frequency is
above the unity-gain frequency of the loop. As such, the signal is simply attenuated by
L 0 before being applied to the quantizer. In the quantizer, the non-linear, tone-generating
mechanisms, occur, similarly as described above6 . However, on the frequency axis, the
picture looks very different:
• the odd harmonics of the input signal are at high frequencies. Moreover, they are
maximally separated from the sample frequency mfs (or harmonics thereof). As
such, aliasing will not cause components in the baseband;
• FM-modulation of the idling frequency directly results in a Bessel component ap-

pearing at baseband. In case of a small input or strong attenuation of the input signal
by L 0 , this is the most important component and its amplitude is almost linear with
the input amplitude;
• the Bessel components generated at high frequencies fold back due to aliasing. This
mechanism becomes important for larger input amplitudes.
Again, all these mechanisms occur in the quantizer. As such, the tones at low-frequencies,
are counteracted by the noise shaping.
In this book, we look for the allowable amplitude of inputs around mfs /2, such that
the amplitude of the first Bessel component does not deteriorate the DR in the wanted
channel. As mentioned, an accurate, quantitative analysis of this problem is not described
in literature, nor do we aim for it here. Instead, this limit is evaluated experimentally
for the conditioning ADCs presented in chapter 9. In addition, we present techniques to
reduce spurious responses for these inputs.
Reduction of spurious responses: Common techniques for reducing spurious re-

sponses aim at de-correlation between the input signal and quantization noise. This can
be achieved by:
• stronger noise shaping; i.e. increasing the order of L 1 ;
• adding “dither”.
The latter technique assumes that the frequency of the dithering signal is larger than that
of the input signal causing the tones. As such, dithering may not be effective for the
high-frequency input that is considered here. In addition, a lot of dithering energy may
be required to lower the tones [83]. Part of the dithering signal appears in the wanted
channel and, consequently, affects the DR.
6 Since the input signal is applied beyond the unity-gain bandwidth of the modulator, this is an ex-
trapolation of the theory for low-frequency inputs. Consequently, this should be considered as a first-order
approximation only.
Alternative to the above, the high-frequency interferer can be attenuated before being
sampled in the quantizer. Possible implementations are:
• a simple (notch) filter with stop-band around mfs /2 preceding the ADC;
• design of L 0 for additional suppression near m fs /2.
Although the first solution is based on preceding, analog signal conditioning the filter can
be much easier than the filter in the full-analog conditioning channel of fig. 5.2. In this
book, the second means, i.e. L 0 -design, is used.
Spurious tones due to analog deficiencies: In practice, spurious tones appearing in

the bandwidth of interest, may as well be due to a very different effect. As such, they may
be caused by:
• parasitic mixing of the input with multiples of mfs /2 (for example due to cross-talk
from divided clocks on the DAC reference, bias lines or supply rails) or with the
high-frequency quantization noise in the output spectrum (for example present on
the substrate, supply or bias lines);
• IM3 -distortion of the input and the feedback signal may occur in the analog blocks.
These problems are common to all analog circuits and can be prevented by linear design
of the input stage, input and feedback paths and by careful layout.
The above discussion does not enable a quantitative prediction of the tones. Still, in prac-
tice the problem of spurious responses is manageable when taking into account the men-
tioned guidelines. This pragmatic approach is in fact similar to the “acceptance” of high-
order designs. Although a solid proof of stability lacks, still, numerous high-order
modulators have been designed successfully or even proven in mass-production [77],
[38], [2], [37], etc. In chapter 8, measurements on spurious responses are included for
various designs.
Intermodulation distortion
As interferers are applied to the ADC care must be taken that intermodulation distor-
tion is sufficiently low. Intermodulation of interferers and/or shaped quantization noise
could corrupt the resolution in the conversion bandwidth of the ADC. Therefore, the
IM3 -performance of the ADC yields another limit on the maximum allowable input
level for interferers. This limit can be calculated from eq 4.12 assuming the transconduc-
tance of the input stage remains the dominant cause of distortion over a large frequency
range.
Obviously, this particular limitation cannot be relieved by the choice of L 0 .

The power/performance analysis of the channel is strongly dependent on the amount of
signal conditioning that happens in the ADC. Therefore, the discussion here is limited
to the basic requirements on the ADC and is detailed later on.
Noise and distortion

The input-referred noise and distortion requirements on the conditioning ADC are
the same as those on the input stage of the full-analog channel and those on the ADC
in the full-digital channel. Hence, the specifications in the last column of table 5.1 apply
for the conditioning ADC (except for p = 1 since, here, the conversion bandwidth
corresponds to the wanted channel only). In an analogy to eq. 5.1, the quiescent current
required to meet the noise and distortion requirements follows:
7/6
IADC ( based CCh) DRi
= Fq G 4/3
MIN
(6.6)
IADC (analog CCh) SINADo
The factor F is an implementation factor used to generalize the expression. All the con-
ditioning ADCs that are discussed in this chapter, target the same performance and obey
the same power/performance relation. Still, their absolute power consumption is differ-
ent. This difference is captured in the factor F. The value of F is determined by the
topology. By definition F = 1 for the feed forward reference ADC, discussed in sec-
tion 6.2; i.e. that ADC is taken as a reference. The ADCs of sections 6.3 and 6.4 have
F = 1.
Sample rate
The advantage of this third architecture compared to the full-digital architecture, is in its
lower sample rate mfs :
• the Nyquist sample rate f s remains as low as in the full-analog channel;
• the over-sample ratio m must be increased though to keep quantization noise low
enough in view of the high DR specification.
As a result, for the reference ADC, the sample rate can be lower by a factor p, when com-
pared to the full-digital architecture. For the conditioning ADCs with a filtering STF
and VGA, the over-sampling can be reduced further. This is shown in sections 6.3 and 6.4.
Circuit bandwidth
Depending on the application, the circuit bandwidth is set either by the sample rate of the
ADC or by the bandwidth of the input signal. The latter may be quite large since
interferer channels can be present over a wide band. This is illustrated in table 6.1 listing
the bandwidth of a wanted channel7 and of the entire band for a few wireless communi-
cation systems. For most examples, it is likely that the sample rate of the ADC is
7 In fact, the channel spacing is listed instead of the channel bandwidth. The channel bandwidth is only
determined by the modulation. The channel spacing also takes into account guard bands in between neighboring
channels.
6.2. SIGNAL CONDITIONING IN THE DECIMATION FILTER 79
Table 6.1: Channel bandwidth and operation band for a few wireless commu-
nication systems
System Channel Band mfs and ref.
GSM 200kHz 25MHz 26MHz in [84]
UMTS 5MHz 60MHz 153.6MHz in [84]
Bluetooth 1MHz 78MHz 64MHz in [32]
IEEE802.11a 20MHz 100MHz 160MHz in [85]
in the same order of magnitude as the operation band. In those cases, the wide range of
interferer channels doesn’t put a significant penalty on the circuit bandwidth.
In the following sections various implementations of a conditioning ADC are

presented. The signal conditioning ranges from “exploiting” the properties of known
topologies in sections 6.2 and 6.3 to explicitly moving an analog filter into the
loop in section 6.4.
6.2 Signal conditioning in the decimation filter

This conditioning channel is analyzed for the example of the feed forward ADC depicted
in fig. 6.6, as a reference case. Generalized, an N th -order loop filter is considered. Here,
N = 3 as an example. Referring to the model of fig. 6.4 L 0 and L 1 can be calculated:
L 0 (s) = −L 1 (s) (6.7)
= a1 H1 (s) + a2 H1 (s)H2 (s) + a3 H1 (s)H2 (s)H3 (s) (6.8)
In this topology, only the error signal is processed in the loop filter. The first filter stage
can have high gain and, by consequence, the following stages are not critical. This results
in a very low-power implementation. Implementing the highest possible gain in the input
stage, is a general advice for a low-power design strategy.
This ADC can be used in the conditioning channel of fig. 6.3.a. In the following,
the tolerance of this topology to interferers is emphasized rather than its performance for
wanted channels. This topology is referred to as the “reference conditioning ADC”,
or, even shorter, as the “reference ADC”.
Figure 6.6: Reference ADC with a 3r d -order feed forward loop filter

In section 6.1.3, the interferer immunity was analyzed for the universal modulator of
fig. 6.4. Here, those results are applied to the reference ADC.
Aliasing
Filling out eq. 6.3, the filter L 0 only has first-order roll-off for inputs near the clock fre-
quency due to the feed forward paths. Typically, the unity-gain frequency of the loop filter
is at 1/6th of the sample rate [25] for stability reasons. Hence, the associated suppression
is rather small (∼ −15dB). The consequent aliasing that occurs in the quantizer is sup-
pressed towards the output by the N th -order noise-shaping of the loop. Hence, the overall
alias-suppression function is of order N +1.
Stable input range
The stable input range (eq. 6.5) of this topology can be derived from the STF. Based on
eq. 6.8 and eq. 6.1 the loop gain (equaling c L 1 for the model of fig. 6.4) and the mag-
nitude of the STF of this ADC are depicted in fig. 6.7.a and b. The exact curve
strongly depends on the actual parameters of the loop. Hence, these graphs should merely
be considered as an example showing some general characteristics. Below the unity gain
frequency f ug of the loop the STF is constant except for the overshoot of several decibels
adjacent to the conversion bandwidth. This is due to the fast phase shift of the open loop
gain around f a ( f a indicates the transition from a high-order to a first-order slope in the
loop gain) and translates into a reduced stable input range for adjacent interferers. Far-off
interferers -beyond the unity-gain of the loop- are suppressed with a first-order slope cor-
responding to the first-order open loop transfer.
Spurious responses
The comments of section 6.1.3 apply.

Figure 6.7: Magnitude of the loop gain and the STF of the reference ADC of
fig. 6.6
Again, the discussion in section 6.1.3 is valid since the dominant non-linearity of this
topology is normally due to the transconductance of the input stage.
The above limitations on the allowable level of interferers are summarized in fig. 6.8
for the 3r d -order topology of fig. 6.6. The four limitations are indicated by a dashed line.
The solid curve indicates the dominant limitation over frequency. Of course, the relative
position of the various curves strongly depends on the actual implementation and this
graph should be considered as an illustration. Still, for most frequencies, the allowable
interferer level of the reference ADC is often limited by the maximum stable input.

As analyzed above, the reference ADC is robust to interferers outside the conversion band-
width of the ADC. By consequence, the conversion bandwidth is designed to accom-
modate the wanted channels only and a bandwidth efficient solution results. Resolution-
wise, though, this efficiency cannot be found. Just as in the multi-channel ADC the
resolution requirements are set by the interferers instead of the wanted signal:
• DR and IM3 are related to the entire input signal including the interferers;
• the stable input range of the ADC is dimensioned to the largest interferer while
the wanted signal can be smaller.
Figure 6.8: Allowable input level for the reference ADC of fig. 6.6
This is illustrated with an example next.
Example:
The input signal of fig. 6.9.a consists of a weak wanted channel and strong interferer
channels. The stable input range of this ADC is dimensioned to accommodate the inter-
ferers. Therefore, the “upper part” of the available DR is not used efficiently. In fig. 6.9.b
the wanted channel is strong. It is converted with a very large SINAD while essentially
much less resolution is required for further processing.
These observations of the “wasted” SINAD or DR indicate that the channel of

fig. 6.3.a does not necessarily yield the best channel architecture in terms of power effi-
ciency. Improved solutions, with signal conditioning in the ADC, are presented in
the following sections.

To a large extent, the generic analysis of section 6.1.4 applies. Some parameters are
detailed here.
The power consumption of the analog and the digital part of the conditioning channel are
-separately- compared to that of the full-analog and the full-digital channel in chapter 5.
Analog part: The generic discussion on the power consumption of a conditioning

ADC (see page 77) fully applies. For noise and distortion reasons, the input stage
consumes as much as any of the other channels. The bandwidth requirements are much
more realistic than those for the full-digital channel, because of the lower sample rate.
Likely, the circuit bandwidth must be dimensioned to the bandwidth of the applied inter-
ferers. In that case, this specification -and the consequent power consumption- is compa-
rably difficult as that of the input stage of the analog channel.
Figure 6.9: Two input scenarios demonstrating the “waste” of SINAD in the
conditioning channel of fig. 6.3.a
Digital part: The digital filter needs to have a larger resolution and run at a higher
sample rate m 3 f s than in the full-analog channel:
3
Idigital ( -channel) m 3 log DR q 2
= 3
(6.9)
Idigital (full-analog) m 1 log SINAD
This discussion is illustrated for the same Bluetooth application as in chapter 5.
Example (continued from page 63:)

Applying eq. 6.6, with the implementation factor F equaling 1, by definition, the current
increase due to the noise and distortion specification remains ∼ 105 . The following
topology is used:
• a conditioning ADC targeting SINAD = 70dB is realized using m 3 = 32, in com-
bination with a complex, feed forward loop filter of order L = 5 (see section 9.2).
Hence, the sample rate becomes 64MHz, which is only 2 times higher than for the con-
ventional channel. Contrary to the multi-channel case, this sample rate is feasible both
for the ADC and for the decimation filter. Filling out eq. 6.9, the power consumption
in the decimation filter increases by a factor of 10 (assuming q equals one) as compared
to the power in the digital part of the full-analog channel.
This example shows that a conditioning ADC can be a very power-efficient al-
ternative to the conventional straightforward digitization of section 5.2.
Figure 6.10: ADC with a 3r d -order feedback loop filter

The very large DR-requirement on the ADC asks for a similar accuracy of the analog
references (i.e. voltage or current reference and clock) used for the feedback DAC. Simi-
larly, the bandwidth requirements of the reference circuits are related to the ADC sample
rate. By consequence, these blocks remain challenging. In fact, this may be one of the
most critical aspects of this particular conditioning channel. For instance, the realization
in section 9.2 needs jitter below 0.01%.
In conclusion, this architecture with signal-conditioning in the decimation filter is

a highly-digitized alternative to the conventional channel. The power consumption of its
ADC is comparable to that of the analog circuits in the full-analog channel. The
decimation filter consumes more than the one in the full-analog channel but this increase
seems acceptable; especially since this contribution shrinks with technology scaling any-
how. The accuracy required of peripheral reference circuits remains stringent. An exam-
ple implementation of this conditioning channel for a Bluetooth receiver is discussed in
chapter 9.
6.3 Signal conditioning with a restricted filtering STF

It is shown that a ADC, based on a loop filter with nested feedback compensation, has
a frequency selective STF. This property can be exploited to provide signal-conditioning
in the ADC. This ADC is referred to as the “feedback ADC”, as a shorthand notation.
Fig. 6.10 shows an example for N = 3. L 1 can be expressed as:
L 1 (s) = − [d3 G 3 (s) + d2 G 2 (s)G 3 (s) + d1 G 1 (s)G 2 (s)G 3 (s)] (6.10)
and realizes the same noise transfer function as the feed forward ADC. Therefore,
this topology achieves the same signal-to-quantization-noise-ratio. Its asset is in a better
immunity to interferers because of the high-order transfer of L 0 :
L 0 (s) = G 1 (s)G 2 (s)G 3 (s) (6.11)
In order to fully exploit this characteristic, variable gain control can be added in the input
stage of the ADC. Then, the feedback ADC truly performs signal conditioning.
First, the interferer immunity is investigated.
6.3. SIGNAL CONDITIONING WITH A RESTRICTED FILTERING STF 85
Figure 6.11: Magnitude of the loop gain and the STF of the feedback ADC
of fig. 6.10

In the feed forward reference ADC the restricting limit on the allowable interferers is due
to the available alias suppression and the stable input range. This is where the feedback
topology excels.
Aliasing
In the feedback ADC, the interferer is subject to N th -order filtering before being
sampled in the quantizer; i.e. the order of L 0 in the nominator of eq. 6.3 equals N at
the interferer frequency. This is different from the first-order filtering in the feed for-
ward ADC. The noise-shaping at the alias frequency remains the same as in the feed
forward case. Hence, the overall alias suppression function is of order 2N (instead of
order N +1 for the feed forward implementation).
Stable input range
The loop gain c L 1 and the STF of the 3r d-order feedback topology of fig. 6.10 are drawn
in fig. 6.11. The frequency f a indicates the transition from a high-order to a first-order
slope in the loop gain; f ug is the unity-gain frequency. Depending on the stability of
the design f a and f ug approximate mfs /20 and mfs /6 respectively [25]. Using eq. 6.1,
Figure 6.12: Allowable input level for the feedback ADC of fig. 6.10
eq. 6.11 and eq. 6.10 the STF of the feedback ADC equals:
Y (s) c G 1 (s)G 2 (s)G 3 (s)
= (6.12)
X (s) 1 + c [d3 G 3 (s) + d2 G 2 (s)G 3 (s) + d1 G 1 (s)G 2 (s)G 3 (s)]
It features 2nd -order filtering for channels between f a and f ug , i.e. in the first-order region
of the loop gain. Beyond f ug the STF shows third-order filtering. Generalized, in case an
N th -order feedback loop filter is used, the STF features filtering of order N −1 for nearby
interferers while far-off interferers experience filtering of order N . The stable input range
is inversely proportional to the STF (eq. 6.5). Notice, the −3dB-frequency of the STF
corresponds to f a . Hence, it can only be lowered at the expense of less noise-shaping. We
will come back on this in section 6.4.
Spurious responses
Spurious responses due to interferers around sub-harmonics of the sample frequency, are
less likely because of the N th -order filtering of L 0 before sampling.
In the feedback topology intermodulation distortion may be higher than what is predicted
by eq. 4.12. Later on, it is analyzed that the signal swing on the various internal nodes
of the feedback topology is much larger than in the feed forward topology. Hence, other
distortion sources may become important as well.
Fig. 6.12 shows the various limitations (dotted lines) on the allowable input level
of interferers over frequency for the example ADC of fig. 6.10. The solid line indicates
the dominant limit. For most frequencies, the restricting limit is due to implementation
aspects (i.e. non-linear circuits) while the topology in itself is largely robust to interferers.
Figure 6.13: Output spectrum of the feedback ADC
Here, the distortion limit is indicated by a flat line, as an example. Depending on the
position of the non-linearity in the loop, the curve may be very different (e.g. the example
in section 9.3).

The increased interferer immunity and the filtering STF of this ADC are optimally
exploited in combination with variable gain control of the input signal8 . A gain setting
between G MIN and G MAX is chosen such that the wanted signal is amplified to the maximum
stable input level. Of course, interferers are amplified likewise. The allowable input level
as characterized in fig. 6.12 should be tailored such that it can optimally accommodate
the input spectrum of fig. 5.1, including the interferers.
Because of the filtering STF and the integrated variable gain control the feedback
ADC performs signal conditioning as depicted in fig. 6.3.b. The signal conditioning
of the merged design is the same as that of the conventional cascade of filter, VGA and
ADC in fig. 5.2. Fig. 6.13 shows the output spectrum of the feedback ADC. Notice
the relative attenuation of the interferers compared to the wanted signal. This feedback
ADC with restricted conditioning is used in a similar way as the feed forward refer-
ence ADC in the sense that:
• the conversion bandwidth only accommodates the wanted signal;
• in the decimation filter, the remaining interferers are attenuated along with the quan-
tization noise.
8 The VGA functionality is easily integrated in the design for instance by making the input resistor
switchable. Since resistors are reasonably linear over a wide bandwidth and voltage range, the VGA can precede
the filtering. By consequence, the input stage (i.e. the first integrator) of the ADC does need to process the
entire bandwidth of the input signal. Only, the DR of the signal is reduced because large wanted input signals are
amplified less than in the full-analog channel. In conventional channels, the VGA normally follows the filter.
Then, the DR and bandwidth requirements are stringent for the filter but become relaxed for the consequent
stages.
Figure 6.14: Two input scenarios demonstrating the efficient use of the available
SINAD
The key difference compared to the reference ADC is in:
• the integrated filtering and variable gain functionality.
The latter characteristic implies the following important consequence:
• while the input-referred DR of the ADC remains large, at the output, the
ADC only needs to provide a moderate SINAD.
It is shown later on that this results in a major overall power saving for the conditio-
ning channel. First, this efficiency is intuitively demonstrated for the same example as in
fig. 6.9
Example:
The same input scenarios as on page 82 are considered. In fig. 6.14.a the small wanted
channel is amplified to the full-scale level of the ADC while the stronger interferer
channels are attenuated to that level because of the filtering. In fig. 6.14.b the gain is
reduced since the wanted signal is strong. The interferers are still attenuated. Because of
the large wanted signal, the ADC noise can be relatively high. Hence, the quiescent
current of the ADC can be adapted to the gain setting (i.e. lowered). By consequence,
the SINAD of the ADC is used efficiently in both input scenarios.
Assuming the same amount of filtering and VGA as in the full-analog channel, this
yields:
• an easy topology: i.e. the same low over-sampling factor, filter order and num-
ber of quantization levels can be used as for the ADC in the full-analog channel;
• an easy decimation filter: i.e. the same decimation filter as in the full-analog channel
can be used with a low ENOB, sample rate and decimation factor;
• relaxed requirements on the DAC voltage, current or time reference: i.e. more noise
and jitter on the DAC is allowed.
The penalty paid for this is in the control loop for the VGA. The complexity of this
loop is comparable to that in the full-analog channel. Another penalty is in the higher
implementation factor. This is discussed next.

It is shown that a feedback topology inherently consumes more power than a feed forward
implementation with the same specifications. On the other hand, the opportunity to inte-
grate variable gain control lowers the average consumption of the ADC and reduces
the specifications on the decimation filter and on the peripheral reference circuits. This
yields an overall power saving.
The feedback ADC

In ADCs with a feed forward loop filter the current consumption of the second and
higher integrators can be low since their noise and distortion is suppressed by the preced-
ing gain. In the feedback topology, the contribution of the consecutive stages cannot be
neglected because the unity gain frequency of the first integrator is much lower than in
the feed forward case. This is due to the two reasons listed below.
• Reverse order of integrator frequencies: Though the feed forward and the feedback
topology realize the same NTF and the same loop gain, the mapping of this function
on the hardware is different. In the feed forward topology, the first integrator H1
has the highest unity-gain frequency. The feed forward term a1 H1 (s) over-rules
the other contributions at high frequencies and provides first-order behavior and
thus (small-signal) stability. On the contrary, in the feedback ADC, this role
is fulfilled by the inner loop d3 G 3 (s). The same reasoning is valid for the other
integrators: for stability reasons the feed forward topology can exploit an ascending
order of unity-gain frequencies while the feedback topology needs a descending
order.
• Larger state-variables: In a feed forward topology only the error signal is fed into
the filter. The error signal primarily consists of the shaped quantization noise. In
case a feedback loop filter is used the entire output spectrum -containing both the
input signal and the quantization noise- is fed back to each internal node of the filter.
By consequence, a strong compensating signal must be provided by each integrator.
This is illustrated in the time domain in fig. 6.15. In order to maintain an accept-
able signal swing at the integrator outputs; i.e. to keep the state-variables within a
Figure 6.15: Internal signal swings in a feed forward and a feedback topology
common magnitude range, the unity-gain frequencies must be scaled down9 . Addi-
tional information (for later use) on the frequency dependence of the internal signal
swings is shown in fig. 6.16. It depicts the magnitude of the transfer function from
the input of the ADC to the output of the integrators based on a linearized model of
the ADCs in fig. 6.6 and fig. 6.10.
Especially for high-order loop-filters the unity-gain frequency of the first integrator(s) is
significantly lower in the feedback topology than it is in the feed forward one. As the
unity-gain frequency of the first integrator drops below the signal bandwidth, noise and
distortion from the second and following integrators contribute increasingly to the overall
SINAD.
Example:
Suppose, Pn (G 2 ) represents a noise source at the input of block G 2 the equivalent input-
referred contribution equals:

Pn,eq,G ( jω) = Pn,G 2 ( jω) (6.13)
2 G ( jω)
1
Beyond the unity-gain frequency of G 1 the contribution of G 2 is amplified and can no

longer be neglected. The same is true for its power consumption.
The above scaling effects and the consequent power increase in the various filter sec-
tions become very cumbersome in high-order filters. This is illustrated in section 9.3.
9 Alternatively, a current-domain filter with a low-impedance load can be used. Since voltage and current are
related via an impedance, a similar reasoning holds. In the current-domain solution, considerable power is often
needed in order to realize a low noise impedance.
Figure 6.16: Linearized transfer function from the input of the ADC to the inter-
nal nodes in a feed forward and a feedback topology

Similar to fig. 5.5 the analysis focuses on the analog and the digital part of the channel
separately.
Analog part: As discussed above, the feedback ADC consumes more than a feed
forward implementation with the same specifications. Hence, eq. 6.6 applies with F > 1.
For instance, it is argued on page 161 that F ∼= 1.5 for that feedback design as compared
to a conventional feed forward implementation.
On the other hand, the integrated, explicit signal conditioning (i.e. filtering of inter-
ferers and VGA) allows for dynamic adaptation of the quiescent current: if the wanted
signal is strong, the input resistance of the ADC is lowered and the quiescent current
can be decreased (see section 9.3). This adaptive biasing lowers the average current con-
sumption. This benefit is hard to quantify since it depends on statistics of the received
signal strength.
The integrated signal conditioning, with strong VGA, results in a moderate SINAD at
the output of the ADC. Hence, the sample rate can be relaxed. Therefore, the required
circuit bandwidth, is probably set by the interferer bandwidth instead of the sample rate.
The requirement is then the same as for the input stage of the full-analog channel.
Digital part: Since the feedback ADC features filtering and VGA, essentially
the same signal as in the full-analog channel is applied to the digital filter. Hence, the
power consumption is as low.

The accuracy required of the references is related to the SINAD at the output of the
ADC, i.e. at the input of the DAC. Since this SINAD is moderate, the accuracy
demand and the power consumption of the reference circuits are relaxed as well. This
is a major benefit of maintaining some analog filtering and VGA, be it merged into the
ADC. The same discussion holds for the bandwidth of these circuits.
In conclusion, this section shows the way to an elegant merge of the filter and VGA
functionality in the ADC. The feedback topology has a worse implementation factor
than the feed forward ADC of the previous channels. Still, because of its excellent
immunity to interferers and the consequent opportunity to integrate VGA, the average cur-
rent can be lowered. In addition, the digital part of the channel and the reference circuits
are as easy as in the full-analog channel. It should also be kept in mind, that the filtering
characteristic of the STF can only be optimized at the expense of the noise-shaping. An
example implementation for a Bluetooth application is discussed in section 9.3.
6.3.4 Conditioning hybrid ADC

For completeness, it is mentioned that the hybrid topology of fig. 6.17 enables a trade-off
between the characteristics of the feed forward and the feedback ADC. It is only
briefly analyzed here. Starting from a feedback topology, feed forward coefficients ci are
added from the ADC input to (some) integrator inputs. Notice that these “feed forward”
coefficients are different from those in the “feed forward ADC”. In the latter case, the
coefficients feed signals from the integrator outputs to the quantizer input. The filter
functions L 0 and L 1 equal:
L 0 (s) = c1 G 1 (s)G 2 (s)G 3 (s) + c2 G 2 (s)G 3 (s) + c3 G 3 (s) (6.14)
L 1 (s) = − [d1 G 1 (s)G 2 (s)G 3 (s) + d2 G 2 (s)G 3 (s) + d3 G 3 (s)] (6.15)
Here, the coefficient ci provide the compensating signal to lower the large low-frequency
signal swing that is otherwise present at the integrator outputs (see fig. 6.15). As the
signal swing drops the unity-gain frequency of the integrators can be scaled up again and
Figure 6.17: Hybrid ADC
Figure 6.18: Magnitude of the STF the hybrid ADC of fig. 6.17 with varying
coefficient c2 and c3 simultaneously
the input-referred noise and distortion from the filter sections G 2 and G 3 -as well as that
of the coefficients c2 and c3 - are reduced proportionally.
On the other hand, each feed forward path reduces the order of the STF by one. The
STF of the ADC of fig 6.17 equals:
Y (s) c3 G 3 (s) + c2 G 2 (s)G 3 (s) + c1 G 1 (s)G 2 (s)G 3 (s)
=c (6.16)
X (s) 1 + c [d3 G 3 (s) + d2 G 2 (s)G 3 (s) + d1 G 1 (s)G 2 (s)G 3 (s)]
In case the coefficients ci equal the coefficients di (with i = 1, 2, 3) a flat STF is obtained
within the unity-gain bandwidth of the loop. As the coefficients c2 and c3 are lowered, the
compensation becomes incomplete and the STF evolves to that of the feedback topology.
Fig. 6.18 shows the STF and the effect of varying the coefficients c2 and c3 jointly. The
same loop gain as in the previous examples is realized. In the general case, the STF of
a feedback ADC with a loop filter of order N , has N poles. In the hybrid topology,
the added coefficients ci introduce Z zeroes in the STF. As such, at high frequencies, the
STF of the hybrid ADC has a slope of N − Z .
The alias-suppression is affected in a similar way: for each coefficient c2 , c3 , etc.,
that is added to the feedback topology, the order of the alias-suppression function is re-
duced by 1. If the coefficients ci equal the coefficients di (with i = 1, 2, 3), the order of
the alias-suppression function becomes N +1, just as for the feed forward ADC.
Hence, the coefficients ci provide some flexibility in tuning the STF without chang-
ing the noise shaping. For a fixed NTF, tweaking these coefficients remains a compro-
mise between the desired STF, the internal signal swings and the scaling of the unity-
gain frequencies. It can be concluded that, in terms of interferer immunity, the hybrid
ADC enables a trade-off between the characteristics of the feed forward and the
feedback ADC. The further analysis is similar to that in the previous sections and
the resulting power estimate reflects the above trade-off. Therefore, it is not discussed in
more detail here.
6.4 Signal conditioning by unrestricted STF design

Instead of compromising between the power efficiency of the feed forward reference
ADC and the robustness to interferers of the feedback ADC, both qualities are united in
a ADC with unrestricted design of a filtering STF, from now on called the “filtering-
feedback ADC” or FFB-ADC. This terminology refers to the fact that at least one
feedback path of the ADC comprises filtering. Notice, this is not the case for the feedback
ADC discussed before: there, the nested feedback paths consist of linear coefficients.
Here, we present two specific implementations of a FFB-ADC. In the remainder, “FFB-
ADC” refers to these examples, although, strictly speaking, it covers a wider range of
implementations.
The presented implementations of the FFB-ADC are based on a conventional ADC
with loop filter H (s) to which a low-pass filter HLPF (s) and a compensating high-pass fil-
ter HHPF (s) are added. Here, a feed forward topology is considered for H (s) in view
of power-efficiency. Fig. 6.19.a and 6.19.b respectively show the parallel and the series
configuration for the added filters (with subscript p to indicate the parallel or s to indicate
the series configuration): Hence:
L 0, p = HL P F, p (s) · H (s) (6.17)
L 0,s = HL P F,s (s) · H (s) (6.18)
The added filters are complementary over the entire frequency range:
HL P F, p (s) + H H P F, p (s) = 1
(6.19)
HL P F,s (s) · H H P F,s (s) = 1
Therefore, the loop gain remains unaltered from that of the conventional ADC (i.e.
the same L 1 -characteristic is realized) and the noise shaping and the stability are the same
as well.
In fact, the same idea can be applied to any of the discussed topologies. The feed
forward ADC is chosen because of its superior power/performance relation. Also, in-
stead of HLPF (s) any filter -or even any function- could have been used in principle. The
only boundary condition is in the fact that the complementary function can be realized
with a certain accuracy. The required matching between the complementary functions is
discussed in section 9.4.
6.4. SIGNAL CONDITIONING BY UNRESTRICTED STF DESIGN 95
Figure 6.19: Reference ADC (a) and ADC with explicit filtering added
in a parallel (b) or a series (c) configuration
The asset of the FFB-ADC, as compared to all previous topologies, is in the combi-
nation of an excellent power/performance ratio and a good immunity to interferers. Note
that the functionality of the FFB-ADC is equivalent to that of a cascade of HLPF and
the reference ADC. Despite of requiring two filters (instead of one in the cascade), the
FFB-ADC yields a more power and area efficient implementation. This is motivated in
section 6.4.3 and illustrated with an example (page 174) in chapter 9.

The FFB-ADC has a better immunity to interferers than the feed forward ADC it is built
on. This is analyzed next.
Aliasing
Suppose the order of HLPF (s) equals M and that of the conventional ADC equals N
then the alias-suppression function is of order M + N +1. It can be calculated that both
for the parallel and for the series configuration the alias suppression improves compared
to the conventional feed forward ADC:
Y ( f ) H (mfs − f )
≈ HLPF (mfs − f ) (6.20)
X (mfs − f ) H ( f )
The improvement corresponds to the amount of low-pass filtering that is added.
Stable input range

The signal transfer function of the parallel and the series configuration are calculated first.
Parallel configuration:
c H(s)
STFp (s) = HLPF,p (s) (6.21)
1 + c [HLPF,p (s) + HHPF,p (s)]H(s)
where c indicates the linearized quantizer gain as introduced on page 72. The DAC gain
is assumed 1 for simplicity. Since the added filters are complementary over the entire
frequency range:
STFp (s) = HLPF,p (s) STFconventional ADC (s) (6.22)
Series configuration:
In an analogy, it can be calculated that:
STFs (s) = HLPF,s (s) STFconventional ADC (s) (6.23)
The STF of the FFB-ADCs equals that of the feed forward ADC (fig. 6.7.b) multi-
plied by the low-pass characteristic of HL P F, p (s) or HL P F,s (s): the ADC explicitly
6.4. SIGNAL CONDITIONING BY UNRESTRICTED STF DESIGN 97
Figure 6.20: Magnitude of the STF of the FFB-ADC of fig. 6.19.a and b (parallel
and series configuration)
provides filtering of the input signal towards the output. Contrary to the previous ADCs
the type, the order and the −3dB-frequency of the filtering can be chosen completely in-
dependent of the loop filter H (s). This is a key asset of the FFB-ADC.
The STF of the parallel and the series configuration are compared in fig. 6.20.a and
b respectively. The same feed forward ADC as that of fig. 6.6 and the following
first-order filter functions are assumed:
a s
HL P F, p (s) = and H H P F, p (s) =
s+a s+a
s+b s+a (6.24)
HL P F,s (s) = and H H P F,s (s) =
s+a s+b
with a < b
The parameter a determines the −3dB-frequency of the low-pass filters. It is chosen to

correspond to the signal bandwidth. The low-pass filter in the parallel configuration is
designed to have a finite gain at high frequencies. This zero is provided by parameter b.
It is chosen to correspond to 3 times the signal bandwidth.
The stable input range is proportional to the inverse of the STF (eq. 6.5) and thus
increases as a function of frequency. It can be tailored by choosing the appropriate low-
pass filter.
Spurious responses
Interferers are suppressed by the cascade of HLPF (s) and H (s) before being sampled.
Hence, they cause less correlated patterns than in the reference ADC.
Depending on the implementation of the FFB-ADC eq. 4.12 applies or other distortion
sources contribute as well. This is further detailed in section 9.4.
As for the feedback ADC, it can be concluded that the topology of the FFB-ADC
is largely immune to interferers. Circuit non-linearities are likely to limit the performance
in most implementations.

The new conditioning channel can realize the same filtering as the feedback ADC
of the previous section but it can even do better. Any filter HLPF (s) can be used as long
as the complementary filter can be implemented with a reasonable accuracy. (It is shown
in section 9.4 that the required matching is not very stringent.) Again, the filtering char-
acteristic is optimally used in combination with variable gain control of the input signal.
Then, signal conditioning corresponding to fig. 6.3.b is provided just as for the feedback
ADC. In addition, the advantages listed on page 88 and the efficiency in terms of DR
(illustrated in fig. 6.14) apply.
A key difference from the feedback ADC of section 6.3 is in an easier and more
flexible implementation. This is discussed next.

It is shown that the “filtering ”-topology hardly consumes more than the feed forward
ADC it is built on. Moreover, the combination with variable gain control of the input
signal reduces the average consumption and relaxes the requirements on the decimation
filter and the peripheral reference circuits.
The ADC
On an architectural level two blocks, i.e. the filters, have been added to the feed forward
ADC. Practical implementations with only a small overhead in power consumption
are possible though.
• Contribution of HLPF : Part of the gain of the first integrator can be shifted into
HLPF (s). (In the parallel configuration, the gain of HHPF (s) then needs to be in-
creased by the same amount.) Alternatively, the order of HLPF (s) and the integrator
sections can be interchanged. This will be demonstrated in section 9.4. Both mea-
sures result in sufficient gain in the input stage such that the noise and distortion of
the following stages can be neglected.
6.5. COMPARISON OF CONDITIONING ADCs 99
• Contribution of HHPF : A simple, i.e. first-order, high-pass filter can easily be inte-
grated in the DAC with passive components only. An alternative solution is given
in section 9.4.
There, an implementation of the FFB-ADC is presented. It achieves F ≈ 1.1 (see page 174).

Again, the analog and digital part of the channel are pair-wise compared to the corre-
sponding parts of the full-analog channel.
Analog part: The implementation factor (with respect to noise and distortion) almost
equals 1 and, thus, is much more favorable than for the feedback ADC or the hybrid
ADC. Because of the integrated signal conditioning, dynamic biasing is possible. It
can be programmed along with the gain setting. Consequently, the average current can be
lower than in the full-analog channel.
Bandwidth-wise, the requirements are comparable to those for the input stage of the
full-analog channel. In this respect, the discussion is the same as that for the feedback
ADC of section 6.3.
Digital part: The discussion on the digital part of the feedback ADC on page 92
is valid here too.

The requirements on the references are as easy as for the full-analog channel and for the
feedback ADC (see discussion on page 92).
In conclusion, the FFB-ADC combines the advantages of the highly-digitized chan-

nels based on the feed forward ADC with signal-conditioning in the decimation filter
(section 6.2) and on the feedback ADC with a restricted filtering STF (section 6.3). It
has a comparable, low power consumption as a full-analog channel of section 5.1. The
filtering STF and the implementation of VGA allow a lower average consumption in the
ADC, an easier decimation filter and relaxed requirements on all references just as for the
feedback ADC. Contrary to the feedback ADC, the filtering STF can be optimized com-
pletely independent of the NTF. An example implementation, again targeting Bluetooth
specifications, is presented in section 9.4.
6.5 Comparison of conditioning ADCs

The various conditioning ADCs are compared in terms of filter topology. This leads to a
better understanding of the constraints and the opportunities of each architecture. Next,
the flexibility and power consumption of the various ADCs is compared in order to distill
guidelines for a power-efficient and flexible design.
6.5.1 Comparison of topologies

Although the presented ADCs all obey the universal model of fig. 6.4, they do differ
in the mapping of the functions L 0 and L 1 on a filter topology. In addition, they may also
realize a different L 0 -function. In view of optimal noise shaping, the function L 1 is the
same for all ADCs, though.
In the comparison, the path from the input of the ADC to the quantizer -described
by L 0 - is referred to as the “forward path”. The path from the output of the ADC to the
input of the quantizer -described by L 1 - is called the “return path”.
Signal conditioning in the decimation filter (section 6.2)

Because of the interferer immunity of continuous-time ADCs, this technique is pos-
sible in combination with any of the discussed ADCs. In case of a feed forward
implementation, this is the single opportunity for signal-conditioning in the “broader”
ADC. In fig. 6.21.a, the universal model of fig. 6.4 is translated to this implementation
topology. All filter sections are in the forward path and only one return path is imple-
mented. Since L 0 = L 1 , the STF is more or less flat over a wide frequency range (eq. 6.1).
It cannot be used for filtering.
Signal conditioning with restricted filtering STF (section 6.3)

These conditioning ADCs are captured by the topology of fig. 6.21.b. Still, all filter
sections are in the forward path. However, multiple forward and return paths are available.
As such, L 0 = L 1 becomes possible, enabling |ST F| < 1 for interferers channels. The
nested feedback paths are essential for obtaining the filtering STF. The forward paths
are optional; e.g. in the hybrid implementation of section 6.3.4 they are introduced for
reducing the internal signal swings. Despite of the additional forward and return paths,
in the feedback and the hybrid implementations, all filter sections of the return path are
shared with the forward path. By consequence, the realizable STF is constrained by the
desired NTF.
Signal conditioning by unrestricted STF design (section 6.4)

This is the single topology where the return path(s) include(s) filter sections that are not
present in the forward path: in fig. 6.21.c, the filters F1 and F2 are added in the return
path. F1 and F2 introduce degrees-of-freedom to design for a filtering STF. Notice, this
is a generalized topology. In the ADCs presented in section 6.4, either F1 is a short or F2
is an open connection. The additional forward path H2 is absent, there.
In conclusion, the nested feedback enables a filtering STF. However, the fact that all
filter sections are shared between the forward and the return path, restricts the achievable
6.5. COMPARISON OF CONDITIONING ADCs 101
Figure 6.21: ADC topology enabling: signal-conditioning in the decimation fil-

ter only (a), signal-conditioning with a restricted filtering STF (b)
and signal-conditioning by unrestricted STF design (c)
STF. On the contrary, adding filter sections in the feedback allows for unrestricted filtering
STF design.
6.5.2 Flexibility
The comparison in this and the following section is summarized in table 6.2. The rating
of the various channels is motivated below.
Flexibility is not considered as a synonym for digitization here. Of course, digitization
helps but analog circuits feature several degrees of flexibility too. For example, the con-
ventional architecture, with a cascade of analog sections, needs many control loops for
tuning of gain, offset or filter frequencies. In order to accommodate the remaining tol-
erances, the circuits need to be designed with sufficient margin. As a consequence, it is
rated rather “inflexible”.
Table 6.2: Comparison of the various channel topologies
Channel P-efficiency Flexibility
full-analog channel (sec. 5.1) 0 −−

full-digital channel (sec. 5.2) −− ++
con.a ADC with:
feed forward reference ADC
(sec. 6.2: con. in the decimation filter) + +
feedback ADC
(sec. 6.3: restricted, filtering STF) - +
FFB-ADC
(sec. 6.4: unrestricted filtering STF) ++ +
a con.=conditioning
The conditioning channels featuring filtering and VGA (i.e. the feedback ADC,
the hybrid ADC and the FFB-ADC) need less accuracy for the analog blocks. Though any
of the ADCs may need tuning of the time-constants in case of large process spread,
the tuning is not critical. It must only guarantee that the noise shaping doesn’t start within
the conversion bandwidth. Furthermore, these channels benefit from the overall feedback
of the ADC: the loop operation reduces the accuracy and performance demands of
the individual filter sections. Hence, the channels are all reasonably flexible in terms
of scaling, robustness, etc.
For some applications, the feed forward ADC, with signal-conditioning in the
decimation filter, may be interesting because it provides the full signal DR to the digital
domain. Obviously, the full-digital architecture offers the highest flexibility. It is the
single architecture that enables full-digital selectivity, both in terms of amplitude and in
terms of frequency.
6.5.3 Power consumption

The conventional analog channel is taken as a reference. As previously discussed, the
conditioning feed forward ADC is likely to consume less: the quiescent current of
its input stage is comparable to that for the first stage of the full-analog channel while the
quiescent current of the following stages can be lower because of the overall feedback.
Likewise, for the feedback, the hybrid and for the FFB-ADC, the quiescent current
of the input stage is comparable as well. However, the overhead of the remaining stages is
different. For the feedback implementation it is largest. For the FFB-ADC, the overhead
is negligible in case the added filters are of low-order. All architectures with a filtering
6.6. CONCLUSIONS 103
STF allow the integration of VGA. Consequently, their quiescent current can be adapted
to the gain setting: if the input signal is large, the bias current can be lowered. This re-
duces the average power consumption.
Finally, the flexibility of the full-digital architecture comes at the penalty of a sig-
nificant power increase. Foremost, the ADC is very difficult and it may run into various
sample rate related limitations (section 3.4). In addition, the accuracy demands on the
analog references become stringent and even the digital filter becomes challenging be-
cause of the high clock frequency.
6.5.4 Guidelines
From the discussion in this chapter and the comparison above, guidelines for the design
of a low-power and flexible conditioning channel are distilled:
• in view of distortion and noise, all channel topologies require about the same quies-
cent current for the input stage. However, the “overhead” current of the consequent
stages can be lowered by applying overall feedback and providing high gain in the
first stage. As such, a conditioning channel based on a ADC -preferably with a
feed forward loop filter- is advised;
• the signal conditioning should happen early in the channel in order to reduce the
bandwidth and resolution requirements in the following of the channel and in the
reference circuits. In the conditioning ADCs, this is achieved by the dedicated
design of L 0 ;
• the sample rate of the ADC must be kept as low as possible in view of power
consumption in the ADC, power consumption in the digital part and accuracy
and bandwidth requirements on the reference circuits;
• digital control -e.g. of the VGA- can be implemented to reduce the average power
consumption of the analog blocks;
• a wise balance between flexibility and power efficiency of the conditioning channel
should be pursued. For example, in the conditioning ADCs presented in this
chapter, the selectivity still happens -to a large extent- in the analog domain but the
burden of tuning and calibration of analog inaccuracies is lifted. This results in a
reasonably flexible, but especially very power-efficient channel.
6.6 Conclusions
In this chapter, the terminology of “conditioning ADCs” and of a “filtering-feedback
ADC” -constituting a specific implementation form of a conditioning ADC- has
been introduced. The ADC is at the heart of an alternative signal conditioning chan-
nel.
An alternative digitization strategy is presented: instead of replacing analog func-

tions by digital processing, these functions can be integrated into a “conditioning
ADC”.
This strategy is slightly less flexible but much more power-efficient than the straight-
forward digitization of section 5.2. Compared to the full-analog channel of section 5.1, it
is likely to result in a lower power consumption for the channel. In addition, it relaxes the
accuracy requirements on the analog reference circuits.
The key pillar of the “conditioning sigma-delta ADC” is a rather under-exploited
characteristic of continuous-time sigma delta A/D converters, namely their inherent im-
munity to interferers.
This immunity enables signal-conditioning in the decimation filter, e.g. using the
feed forward reference ADC, or even in the ADC, e.g. using an ADC with nested feed-
back in the loop filter. As an improvement on the latter, restricted and power in-efficient
form of signal conditioning, the FFB-ADC is presented. It features excellent power effi-
ciency in combination with explicit and flexible conditioning.
This is illustrated with example implementations in chapter 9.

Chapter 7
Digitization of the inter-die

interface
In the previous chapter, alternative solutions for the digitization of the conditioning chan-
nel have been presented. In practice, the conditioning channel may extend over multiple
dies, possibly in a different technology. The ADC can be integrated with the analog part
or it is put on the digital die. Depending on this partitioning choice, the inter-die interface
will be digital or analog. It is shown in this chapter that the digitization of the condition-
ing channel leads to a digitization of the inter-die interface as well. This is depicted in
fig. 7.1.
The chapter starts with some general considerations on the choice of the inter-
face. Next, possible IC partitioning scenarios for a based conditioning channel are
compared with respect to power consumption in the associated interface. The analysis
on power consumption supports the on-going digitization of the inter-die interface for
high-resolution conditioning channels.
7.1 Considerations
In case of a multi-die solution, the choice for an analog or a digital interface can be
determined by a number of factors, a.o.:
• technology choice for ADC/cost: The ADC can either be integrated in an advanced
CMOS technology or a technology with high-performance analog capabilities. Of-
ten, the choice for one or the other is cost-driven.
• electro-magnetic interference: In case of a digital interface the communication can
be a source of electro-magnetic interference affecting the performance of other cir-
cuits. The opposite is true for an analog interface: it is susceptible to interference
from other sources. A differential implementation is therefore preferable in all
cases. For a digital interface low-swing operation is favored while for an analog
link a large signal is preferred.
105
106 DIGITIZATION OF THE INTER-DIE INTERFACE
Figure 7.1: Analog interface for a full-analog conditioning channel (a) and
digital interface for a full-digital channel
• pin count: An analog, differential interface requires two pins per signal path. The
same is true for a digital, differential interface. Likely, though, it needs many more
signal paths because of the number of parallel output bits. These bits can be serial-
ized to reduce the pin count. Next to the serializing/de-serializing this also requires
one additional signal path to transmit a synchronization signal. In addition, a digital
interface requires transmission of the clock, unless clock recovery is implemented.
• standardization: In an industrial realization, compatibility of the analog die with

various digital ICs and vice versa is of major importance. As a consequence, the
inter-die interface is being standardized for various applications. This argument is
illustrated for the example of a GSM chip set.
Partitioning of a GSM chip set: Fig. 7.2.a shows a typical chip partitioning for the
GSM chip-sets of the 1990’s: it has an analog interface between the RF IC and the
digital baseband with the ADCs and DACs. Meanwhile, a consortium of companies
is defining a digital interface standard “DigRF” for the next generation of GSM
solutions (see fig. 7.2.b and [86]). Similarly, standards like a.o.‘MiPi” [87], “JC61”
[88] and “BlueRF” are being developed.
• power/performance of the interface: A last argument for choosing a digital or an

analog interface is in their different power/performance relation.
This is analyzed next. The analysis of the interface is conducted from the angle of digiti-
zation of the channel. Hence, it is of a limited nature.
7.2 Power in the interface

In a based conditioning channel the IC partitioning can be such that the inter-die
interface is either analog (before the ADC), digital after decimation or digital but before
decimation (see fig. 7.3). The power associated with these three possible interfaces is
7.2. POWER IN THE INTERFACE 107
Figure 7.2: Conventional GSM solution with analog inter-die interface (a) and
proposed digital interface standard “DigRF” (b) for next genera-
tion, highly-digitized GSM solutions [86]
Figure 7.3: Possible IC partitioning for based conditioning channels

discussed next. It is based on the power calculations of analog and digital data interfaces
in appendix E and in [89].
7.2.1 Analog interface

From the calculations in appendix E the power consumption of an analog interface fol-
lows:
P ∼ 22ENOB BW (7.1)
The result is similar to the power/performance relation defined for analog conditioning
circuits in chapter 4.
7.2.2 Digital interface after decimation

The calculations in appendix E distinguish between a full-swing and a low-swing or slew-
rate controlled interface.
The dissipation in a full-swing interface is dynamic only, i.e. it is solely due to the
charging and discharging of a load capacitance. Simplifying eq. E.7 the power consump-
tion in a full-swing interface follows:
P ∼ ENOB fsw (7.2)
where frequency f sw is defined as the inverse of the average number of 0 − 1 transitions

in the data. It is considered proportional to the sample frequency. For simplicity, it is
assumed that all bits are “effective” and, thus, ENOB is filled out. For decimated data, the
difference between the number of bits and the effective number of bits is small.
Alternatively, a low-swing interface can be implemented for interference reasons. In
that case, part of the dissipation is dynamic and follows eq. 7.2. Another part is static: it
is due to the DC biasing current and is independent of ENOB and f sw .
Hence, it can be concluded that the current consumption of a digital interface is

proportional (in case of a full-swing interface) or less than proportional (in case of a
low-swing interface) to the data rate, i.e. the product of ENOB and f sw .
7.2.3 Digital interface before decimation

A ADC exchanges resolution in amplitude (i.e. ENOB) for resolution in time (i.e.
f sw ). This exchange is not linear. By consequence, a ADC has a higher output data
rate than what is essentially required in view of the ENOB and the signal bandwidth; i.e.
not all output bits are “effective”. This is true for encoding in general, but especially
for single-bit encoding:
• single-bit quantization adds a considerable amount of quantization noise (much

more than high-resolution Nyquist A/D conversion)
7.2. POWER IN THE INTERFACE 109
Figure 7.4: Required over-sampling ratio m as a function of the ENOB target

and the effect of increasing the filter order L
• single-bit encoding reduces the maximum modulation depth of the digital out-
put by a factor of two compared to the digital full-scale power.
In case the IC partitioning is such that the ADC and the decimation filter are on separate
dies, power is “wasted” for the communication of quantization noise in between the dies.
In order to express the data rate at the output of the ADC in terms of ENOB and
signal bandwidth, eq. 3.1 is used. The following relation between the over-sample ratio
m and the effective number of bits ENOB can be derived:
π
m∼
ENOB
=2 L 1
(7.3)
(2L + 1) 2L+1
with L indicating the order of the loop filter. This suggests that the required over-sampling
increases exponentially as the ENOB-target grows. In practice, though, m is a much
weaker function of ENOB, because:
• L is a function of m: as m increases, a higher-order loop filter is normally imple-

mented as well
1
• the function (2L + 1) 2L+1 only varies between 1 and e1/e for positive values of L.
Hence, this variation can be neglected.
The above is clarified with the curves in fig. 7.4. The plotted relation is based on C-
simulations of a bank of available filters for various values of m. The combination yield-
ing the highest ENOB is plotted. Alternatively, eq. 3.1 could have been used but, as
mentioned on page 23, that estimate may be too optimistic because it doesn’t take into
account stability related limitations.
The switching frequency of the digital link is proportional to m. Filling out relation
eq. 7.3 in eq. 7.2 the dynamic dissipation of a digital inter-die interface with modulated
data becomes:
ENOB
P∼2 L BW (7.4)
Finally, note that an important asset of a digital link with bitstream data is in its robustness
for bit-errors in the code. Contrary to a multi-bit code, in the bitstream code all bits are of
equal weight and any bit-flip causes only a small error. This remark links to the historical
overview as described in section 3.1.
7.2.4 Comparison
For the three interfaces discussed above, the current consumption versus ENOB is plotted
in fig. 7.5. Since the above relations only predict a proportionality, the absolute position
of the curves depends on the implementation, the various specifications (for example the
load capacitance, the signal swing, etc.) and the margins taken in view of interference.
Here, eq. 7.1 and 7.2 are normalized such that these graphs intersect at ENOB = 10 as
an example. It is observed that, for low-resolution -probably large bandwidth- applica-
tions an analog interface consumes least while in case of a high-resolution -probably low
bandwidth- application a digital interface is favorable.
The curve of eq. 7.4 (corresponding to the encoded data) is fixed relative to that
of 7.2 (corresponding to the decimated date). As such, the intersection of eq. 7.1 with
eq. 7.4 lies at a higher ENOB-value: assuming both types of digital interfaces are imple-
mented on the same hardware, the interface before decimation consumes more than the
one after decimation because of the encoded data. The difference is less dramatic
though than what is suggested on first sight when comparing eq. 7.2 and eq. 7.4 (see dis-
cussion on m as a function of ENOB). For high-resolution channels, the interface before
decimation remains attractive compared to the analog interface.
Depending on the application, the decimation filter may also include scaling of
the word-length (i.e. digital VGA: the input-referred DR can be reduced to the SINAD
required for further processing). In that case, a digital interface with decimated data be-
comes significantly more attractive because of the further reduction in data rate.
7.3 Application to the conditioning channels

The above findings are applied to the channels discussed in chapter 5 and 6. A deci-
sive conclusion on the choice for the interface should include the application (for exam-
ple high- or low-resolution) and the considerations listed on page 105. Here, a high-
resolution, moderate bandwidth application, as for example wireless communications, is
assumed and power consumption is the dominant decision criterion.
Under these assumptions, the following guidelines are distilled.
• Choose an analog interface for the conventional channel of section 5.1: since a
lot of analog conditioning is applied only a limited DR is required and an analog
interface consumes least.
Figure 7.5: Current consumption in the inter-die interface as a function of the

number-of-bits (with arbitrary intersection for ENOB = 10)
• Choose a digital interface with decimated data for the multi-channel architecture
of section 5.2: this is the best choice for a high-resolution link. Moreover, the
decimation filter may include channel selection and reduce the signal DR to the
SINAD needed for digital processing.
• Choose any digital interface for the conditioning architectures: since the ADC
in these architectures includes some signal conditioning its output data rate is mod-
erate and a digital link remains favorable. Because of the robustness of -
encoded data to bit-flips, the digital interface before decimation may be preferred..
In view of power consumption, the general advice is to perform the signal conditioning
before interfacing to a following die. By consequence, as the signal conditioning is being
digitized, the interface becomes digital as well.
7.4 Conclusions
For low-resolution, large bandwidth applications, an analog inter-die interface is preferred
for power reasons. For high-resolution, low bandwidth applications, a digital interface
consumes less.
In an analogy to the results for signal conditioning (chapter 4), it is found that the
power consumption of an analog data interface is exponential in ENOB. Power consump-
tion of a digital interface is linear or less than linear in ENOB, at least, in case decimated
data is assumed. In case the data is encoded, the power consumption of the digital
interface shows a weak-exponential dependence on ENOB.
For highly digitized conditioning channels, a digital inter-die interface should be
used in view of power consumption.
Chapter 8
Highly analog and highly digital

channels for FM/AM radio
This chapter illustrates the theory in chapter 5 on highly analog and highly digital con-
ditioning channels. A dual-mode conditioning channel for use in an FM/AM car radio
receiver is presented. Elaborating on the rest of the book, this channel includes some IF
blocks in addition to the baseband functionality. In FM mode, the signal conditioning is
highly analog (as described in section 5.1). In AM mode, the same conditioning channel
is re-used. As such, a sufficiently large bandwidth becomes available for multi-channel
A/D conversion (as described in section 5.2). However, the multi-channel aspect leads
to very challenging linearity requirements. This becomes obvious from the design of a
CMOS VGA around the IF frequency and from the design of the ADC performing the
multi-channel A/D conversion.1
8.1 System
Although FM and AM radios have been around for decades, there is a clear need for
further integration and digitization of the system. This is especially true for a car radio
system. Both cost- and performance-wise this is a very demanding application. Digitiza-
tion of signal processing and conditioning brings the following advantages a.o. [90]:
• higher quality: DSP algorithms can be used to reduce the effect of a.o. ignition
noise and interferers. Software-controlled adaptive bandwidth of digital channel-
select filters improves the signal quality and fidelity: this bandwidth can be chosen
depending on reception conditions such as the strength of neighboring channels;
1 This chapter heavily relies on work by Eric van der Zwan. He derived the specifications on the ADC and
defined the ADC architecture. He also took the lead in the publication of [2], of which some parts have been
re-used here. The implementation of the ADC -though heavily inspired by previous work of Eric a.o. [7]-
was performed by the current author. In addition, the analysis, the evaluation in the present context and the
benchmark are original work, just as all VGA-work.
113
114 HIGHLY ANALOG AND HIGHLY DIGITAL CHANNELS
Figure 8.1: FM/AM radio with analog demodulation
• more features: Many aspects of the radio functionality and performance can be
parameterized and can be made user-controllable;
• more flexibility: Quality and features can be tailored to any market or customer
because of programmable, digital filtering and processing.
Below, a radio with analog signal conditioning and demodulation is described first. Next,
a radio with digital demodulation is presented. In FM-mode, the signal conditioning
remains highly analog but in AM-mode, multi-channel A/D conversion and full-digital
selectivity are implemented.
8.1.1 Conventional radio with analog demodulation

Fig. 8.1 shows a block diagram of a FM/AM radio receiver with analog demodulation.
Depending on the continent, FM radio operates in between 65-108MHz. AM channels,
including LW, MW and SW frequency bands range from 145kHz to 10MHz. The radio
front-end converts the FM or AM antenna signal to an intermediate frequency (IF) of
10.7MHz. Analog FM demodulation takes place after 200kHz channel selection at the IF.
The channel-selection may consist of more than one external filter; e.g. the high-end car-
radio application described in [91] needs four external filters. Typically, these are high-
quality, surface acoustic wave filters (SAW) possibly combined with a coil. AM signals
are processed by a double conversion receiver using the same 10.7MHz frequency as first
IF. After mixing to a second IF (for example 450kHz), 9kHz channel filtering (by another
expensive, external filter) and amplification by a VGA, analog AM demodulation takes
place. The demodulated FM signal or the AM audio signal is digitized in the audio signal
processor IC. This signal processor may provide functions like interference absorption,
stereo decoding, radio data system (RDS) decoding and audio controls such as volume,
balance, tone control and dynamics compression. At the output, the digital signals are
converted to analog signals, which connect to the power amplifiers.
Notice that this radio receiver has full-analog selectivity: even the demodulation of
AM and FM signals happens in the analog domain. On the other hand, a lot of digital
8.1. SYSTEM 115
Figure 8.2: FM/AM radio with digital demodulation
Figure 8.3: Input spectrum to VGA and ADC in FM (a) and AM (b) mode
signal processing is already applied in car radios. Hence, this is a further stimulus to
digitize the receiver as well.
8.1.2 Radio with digital demodulation

A FM/AM receiver with a higher level of digitization (a.o. digital demodulation) is shown
in fig. 8.2. In this solution, demodulation of the radio signal is performed digitally. An
FM SAW channel filter, usually about 200kHz wide and with ∼40dB of adjacent channel
suppression, is still required to protect the VGA and the ADC from strong interferer chan-
nels. Its specification is however relaxed compared to that in fig. 8.1 because it doesn’t
have to provide the full channel selectivity: final channel filtering takes place in the digital
domain. The A/D conversion happens at the 10.7MHz IF, both in FM and in AM mode.
In fact, it is detailed later on that this IF-ADC consists of a low-pass ADC with a
passive mixer integrated in its input stage [38].
Highly-analog signal conditioning in FM mode: In FM mode a single 200kHz chan-

nel is converted to digital. The neighboring interferer channels are not completely re-
moved but they are sufficiently suppressed in order not to influence the ADC design (see
fig. 8.3.a). Hence, in FM mode, the signal conditioning is a compromise between that of
the full-analog channel of section 5.1 and that of the channel in section 6.2 with signal-
conditioning in the decimation filter.
Highly-digital signal conditioning and multi-channel A/D conversion in AM mode:

Compared to the analog receiver of fig 8.1, the external AM channel filter can be omit-
ted. This reduces the required PCB area, the number of pins, power consumption in I/O
Figure 8.4: Technology partitioning with bipolar (a) and with CMOS (b) VGA
drivers, interference, etc. AM channel selection is shifted to the digital domain com-
pletely: since the AM channels are only 9kHz wide, about 20 channels pass through the
200kHz FM filter (see fig. 8.3.b)). Hence, in AM mode, this is a highly-digital conditio-
ning channel with multi-channel A/D conversion just as in the channel of section 5.2.
8.2 VGA design

The main research focus of the car radio concept of fig. 8.2 is in the design of the ADC.
Still, here, the analog conditioning in the VGA preceding the ADC is presented as well.
The original technology partitioning is depicted in fig. 8.4.a: the RF front-end and the
VGA are on a bipolar die, the IF-ADC is integrated with the digital processing and the
other mixed-signal circuits in a 0.25µm-CMOS technology. In an attempt to alleviate the
design of the bipolar VGA, the design of a CMOS VGA is explored (fig. 8.4.b). The key
challenge of the design is in a very high linearity (i.e. IM3 < −80dBc in AM mode) at a
differential output swing of 1.4V pp at a supply of only 2.5V.
Integrating the VGA in CMOS has obvious benefits with respect to pin-count, re-
liability and compatibility as the gain control loop can then be closed on one die. The
fixed pre-amplification is kept on the bipolar IC in order to maintain noise immunity at
the interface between both ICs. The interface between the ICs is differential. The system
furthermore requires four gain steps of about 6dB each and a differential output swing
of 1.4V pp from the VGA to the 62.5kOhm input impedance of the ADC (see sec-
tion 8.3).
8.2. VGA DESIGN 117
Figure 8.5: Highly linear VGA with feed back
8.2.1 Highly linear VGA design

A highly-linear VGA is depicted in fig. 8.5. Its basic operation and an AC analysis are
discussed next.
Basic operation
Fig. 8.5 depicts a two-stage amplifier with negative feedback where the input signal is
applied to the positive input terminal [92]. The current sources force the input transistor
MP1 to conduct a fixed quiescent current. Thus MP1 operates as a source follower and
copies the input signal at its gate over resistance R1 , being a highly linear conversion
impedance. The second amplifying stage acts as a class-A amplifier. Transistor MN1 pro-
vides the signal current. It may be very non-linear, as its distortion is reduced by the gain
in the preceding stage. Because of the feedback, transistor MP1 only needs to be linear
with respect to a small error current i instead of the entire signal current i. The larger
the loop gain at the IF frequency, the smaller the error current and the less linearity is
required of MP1 . Thus, the presented VGA topology achieves low distortion by exploit-
ing the gain-bandwidth product available in advanced CMOS-technologies instead of the
(decreasing) inherent linearity of the devices.
P-MOS devices have been used for the input pair of the amplifier. This allows con-
trolling the back-gate voltage: a bias circuit for the Nwell of the input transistors prevents
back-gate modulation. At the same time, this prevents bandwidth limitations or linearity
problems due to the Nwell capacitance. The bias circuit consists of a replica of the input
branches that copies the AC input signal to the Nwell of the input P-MOS transistors. By
consequence, the charging current of these nodes does not interfere in the signal path.
The input common-mode voltage is controlled by a feed back loop that is outside the
present scope. The output common-mode voltage is set by the ADC that follows this
VGA. All four branches conduct 500µA. This quiescent current is needed to meet the
noise and bandwidth requirements.
The closed-loop gain of the VGA approximately equals 1+ R2 /R1 . It is shown in
eq. 8.4 that this approximation still holds around the 10.7MHz input frequency. The gain
variation is implemented by varying R1 . Because of the chosen resistances and including
the on-resistance of the switches, gain settings of 5.8dB, 12.5dB and 17.0dB are obtained.
(The requirement for a fourth gain setting (see introduction) is met when by-passing the
VGA and thus realizing 0dB of gain.) Switching R1 instead of R2 has an advantage in
terms of loop stability (see below) and for linearity. Details on the latter are described
in [93], as well as comments on the accuracy of the gain setting.
Loop analysis
For this analysis, we refer to the universal amplifier model of fig. 8.6.a and to the para-
meters defined in the single-ended schematic of fig. 8.6.b. The open loop gain A, the feed
back factor k and the input factor d of the VGA can be calculated:
gm1 gm 2 (R2 + R1 // gm1 )

1
A= (8.1)
sC1 1 + sC2 (R2 + R1 // g1 )
m1
R1 // gm1
1
k= (8.2)
R2 + R1 // gm1
1
1
gm1
d= (8.3)
1
gm1 + R1 //R2
The loop gain equals Ak and the stability can be analyzed by solving 1+Ak = 0. The
poles and zeroes of the circuit are indicated in fig. 8.6.b. In order to simplify the expres-
sions, the high-frequency pole-zero pair at node N3, is left out of the above equations.
In fact, fig. 8.6 shows a simplified schematic. In the actual circuit, Miller compensa-
tion is applied to split the poles at nodes N1 and N2 in fig. 8.6.b and to guarantee stability.
In fact, the stability is conditional since the second pole depends on the load impedance.
In the present application, though, the load of the VGA is set by the input impedance of
the ADC that is discussed in section 8.3. It represents a resistive load of 62.5k.
This impedance can be neglected compared to the load of the feed back path.
Finally, it is mentioned that the gain-bandwidth product of the loop exceeds 30GHz
in all modes. This guarantees fast settling in case of a transition to a different gain setting.
Assuming strong degeneration (i.e. gm1 R1 1), the closed-loop gain can be ex-
8.2. VGA DESIGN 119
Figure 8.6: Universal amplifier model (a) and single-ended VGA schematic (b)
for analyzing the loop
Figure 8.7: Closed-loop transfer of the VGA (simulated)

Figure 8.8: Measurement set-up
pressed as:
dA R2 gm2 1
= 1+ (8.4)
1 + kA R1 gm2 + sC1 1 + sR2 C2
Its magnitude and phase are plotted in fig. 8.7 for the three gain settings of the VGA.
Hence, the closed-loop gain of the VGA approximately equals 1+ R2 /R1 for frequencies
within the closed-loop bandwidth determined by f −3d B = gm2 /C1 . In a first-order approx-
imation, the closed-loop bandwidth is the same for all gain settings ( f −3d B ∼ 100M H z)
since it is independent of R1 . Similarly, the loop stability is independent of R1 as well.
On the contrary, any gain setting approach varying R2 would interfere with loop stability:
i.e. it would shift pole at node N 2 in fig. 8.6.b.
8.2.2 Evaluation
The high linearity target requires a dedicated measurement set-up to evaluate the VGA.
This set-up is discussed first. Next, the measurement results are listed and evaluated in
view of the radio requirements, of the simulation results and of the power/performance
relations derived in chapter 4. Finally, the overall VGA performance is bench-marked
with state-of-the-art published designs.
Measurement set-up
The VGA has been processed as a stand-alone circuit with an active area of 0.038mm 2
in 0.25µm CMOS technology. The measurement set-up is drawn in fig. 8.8. A highly-
linear, multi-source generator allows two-tone intermodulation tests around the 10.7MHz
IF frequency. Using a transformer, a differential input signal is applied to the VGA.
The output of the VGA is measured using the IF-ADC that is discussed in the next
section. This ADC has integrated quadrature mixing and, by consequence, provides a
complex output signal around a low-IF frequency. It was fully characterized before-hand.
The VGA and the ADC are on separate dies but have been packaged together in a multi-
die module (see fig. 8.9). The bitstream output of the ADC is captured by a logic ana-
8.2. VGA DESIGN 121
Figure 8.9: Micrograph of multi-die module combining VGA (die1) and IF-
ADC (die2)
lyzer. Further analysis can then be conducted in the digital domain and at the baseband
frequency. This set-up is chosen because of the stringent linearity requirements.
Measurement results
Fig. 8.10 shows the output spectrum in case of a two-tone measurement at an intermedi-
ate gain of 12.5dB. Two tones -equal in amplitude- are applied such that they add up to
a 1.4V pp output signal for the VGA. Since this amplitude corresponds to the full-scale
input level of the ADC, each tone appears at −9dBFS at the digital output. The 3r d order
intermodulation components fall into the band of interest and are at a level of −71dBc , i.e.
71dB below the carrier tones. From the evaluation of the ADCs, it can be concluded that
the measured IM3 -distortion is dominated by the VGA non-linearity. The other spectral
tones (image frequencies and DC offset) are generated by the IF-ADC with quadrature
mixing.
The noise floor in fig. 8.10 combines the contributions of the VGA and the ADC.
VGA and ADC separately, have an output-referred DR of 96 and 97dB respectively (mea-
sured in 9kHz bandwidth for an AM channel). Taking into account the gain range of the
VGA the entire CMOS IF receive path achieves an input-referred DR of 110.5dB in a
9kHz AM channel. The output-referred noise density of the VGA is 84nVrms/sqrt(Hz).
It is dominated by the noise of the current sources and, thus, is largely independent of the
gain settings. The noise density rises for frequencies larger than 300kHz due to the noise
shaping of the ADC.
Figure 8.10: Output spectrum (after quadrature mixing nd A/D conversion)

for an IM3 test
Table 8.1 summarizes the major performance parameters of the VGA. Because the
VGA architecture is based on feedback its performance is robust with respect to variations
in supply voltage, quiescent current, processing, etc. Moreover, it is very well scalable to
new CMOS technologies.
Discussion
The presented CMOS VGA provides a large output signal swing at IF while keeping
IM3 -distortion low (see benchmark in the following section). Still, the linearity is not
good enough for the AM radio application because of the multi-channel aspect. A po-
tential solution is to scale down the signal swing in the channel: IM3 -distortion then
decreases quadratically. Alternatively, the bipolar VGA (at a higher supply voltage) is
used or the VGA can be integrated into the ADC. The latter solution is discussed for
the ADCs targeting Bluetooth specifications (sections 9.3 and 9.3).
The measured distortion is higher than what is predicted from simulations. This can
be due to a.o. imperfections in the device modeling, numerical problems in the distortion
simulation or artifacts in the circuit modeling or the measurement set-up. The discrepancy
was investigated in various tests.
• Effect of decreasing the signal swing: The IM3 -distortion drops quadratically.
Hence, the circuit is not in compression and the distortion is due to a weak non-
linearity.
• Effect of varying the supply voltage: The IM3 -distortion is largely independent of
the supply voltage.
8.2. VGA DESIGN 123
Table 8.1: Measured VGA performance
differential output range 1.4V pp

gain settings 5.8d B/ 12.5d B/ 17.0d B
IM3 a < −72dBc / < −71dBc / < −67dBc
√
output noise density < 84nVrms / Hz
−3dB-bandwidth 100MHz
Iquiescent (VGA + bias circuit) (2.1+0.6)mA @2.5V
active area 0.038mm 2
technology 0.25µm-CMOS, 1P, 6Al
a at 10.7MHz and 1.4V

pp output
• Effect of varying the quiescent current: In the highest gain settings, the IM3 -
distortion increases slightly if the quiescent current is decreased below its nominal
value. For the setting with gain 5.8dB, the IM3 -distortion is independent of the qui-
escent current. This is also the case -for all gain settings- if the quiescent current is
increased beyond the nominal value.
The latter experiment is illustrated in the graph of fig. 8.11. In the setting with 17.0dB or
12.5dB of gain and with a quiescent current below the nominal current, the IM3 -distortion
seems to follow a slope of −1 or −2/3 as a function of the quiescent current. This rela-
tion is expected for the non-linearity of a transconductor or a degenerated transconductor
(eq. 4.7), respectively. In this bias region, and for these gain settings the input transcon-
ductor of MN1 is only weakly degenerated. Hence, it may cause the dominant distortion
here.
Beyond the nominal quiescent current or for a gain of 5.8dB, the distortion slowly
rises. However, the distortion remains low over a large variation in the quiescent current.
Therefore, it may be due to a dominant non-linearity outside of the actual VGA loop.
The curves in fig. 8.11 also prove that the VGA is biased in a region with an optimal
power/performance ratio; i.e. this is a local minimum in terms of distortion.
Benchmark
The presented VGA does not meet the very challenging IM3 requirements set by the multi-
channel aspect in AM mode. Still, compared to published CMOS VGAs, it performs very
well. This demonstrated by the comparison to state-of-the-art published designs with
similar bandwidth and gain targets in table 8.2. All these designs use local or global feed
back. By consequence, the power/performance relation of eq. 4.7 with α = 2/3 is valid
(assuming a transconductor limits the performance) and the performance of these designs
Figure 8.11: IM3 -distortion as a function of the quiescent current (normalized

to the nominal quiescent current)
can be compared using the FOM of eq. 4.15. Although this FOM has been derived for
a ADC, it is equally applicable to analog circuits using feedback (see discussion on
page 52).
This comparison proves the excellent power/performance of the presented VGA.
[95] achieves even lower distortion but benefits from a higher supply voltage (reducing
a.o. the non-linearity of the switches).
8.3 ADC design

Depending on the gain range of the VGA, the DR at the input of the ADC is about 70dB
over 200kHz bandwidth in FM mode. In AM mode, again depending on the VGA range,
the required DR at the ADC input may be larger than 90dB in 9kHz. This means that in
AM mode a lower noise density must be achieved than in FM mode. The multi-channel
aspect also results in severe requirements in terms of linearity and spurious-free DR:
distortion may introduce intermodulation or other spurious components into the wanted
channel. This makes the FM/AM ADC more difficult to realize in AM than in FM mode.
8.3.1 Conventional solutions

Digitization in fig. 8.2 may be performed by a wide-band ADC [90]. A resolution of 10
bits in 11MHz bandwidth may be sufficient to achieve the noise specifications. However,
the linearity requirements for AM are difficult to meet at the 10.7MHz input frequency.
Therefore, an analog AM channel filter is still necessary, which is applied after mixing to
a second IF of 450kHz.
8.3. ADC DESIGN 125
Table 8.2: Comparison of VGA performance
Reference [94] [95] This work [93]
gain range −2 ∼ 12dB 0 ∼ 19dB 5.6 ∼ 17.0dB

f −3d B 15MHz 125MHz 100MHz
output swing 1V pp 2V pp 1.4V pp
IM3 a −56.5dBc −74dBc −67dBc
√ √ √
output noise density 67nVrms / Hz 55nVrms / Hz 84nVrms / Hz
current consumption 4mA 6.4mA 2.1mA
supply voltage 5V 3.3V 2.5V
technology 0.5µm-CMOS 0.35µm-CMOS 0.25µm-CMOS
FOM of eq. 4.15 1 × 10−18 J 1 × 10−19 J 2 × 10−19 J
a for inputs around 10MHz and at maximum gain
Alternatively, bandpass modulation may be used [96], [97], [98], [99], [100],
[101] and [102], providing the required resolution in the bandwidth of interest only. So-
lutions for AM [97] and for FM [98], [99] have been shown but with limited performance
and relatively high power consumption. Further development resulted in combined so-
lutions for AM and FM [96], [100]. Again, the input stage has to be very linear at the
10.7MHz IF to prevent intermodulation of neighboring channels in AM mode. This is
very difficult.
A different approach is to use a phase-locked loop (PLL) to digitize and demod-
ulate the FM radio signal directly after the limiter in fig.2 of [103].
Because of the high DR of the radio signal and the intermodulation distortion re-
quirements, none of these techniques eliminates the need for a high-Q channel filter.
8.3.2 ADC with integrated passive mixer

The ADC that is presented here mixes the 10.7MHz input signal to a second low inter-
mediate frequency, and uses low-pass, continuous-time modulation to digitize the
signal. To this purpose, a passive mixer is integrated into the input stage. This technique
was first presented in [38]. Here, it is briefly reviewed together with the global frequency
planning of the IC. First, the choice for a low-IF topology is motivated and the consequent
requirements on image rejection are discussed.
Low-IF topology: The second mixer may convert the wanted channel to DC (“zero-
IF”) or to an offset frequency (“low-IF” or “near-zero IF”). Both methods have their ad-
vantages and disadvantages [104]. Using zero-IF, the required ADC bandwidth is as low
as possible. The main disadvantage is the sensitivity to flicker noise and DC offset. The
latter may be introduced by circuit mismatch, oscillator self-mixing or second-order dis-
tortion in the mixer. Flicker noise as well as DC offset are located in the middle of the
wanted channel and therefore cannot be removed without loss of part of the signal. This
is the main reason why mixing to a low IF is preferred in this application: the low IF fre-
quency is chosen such that flicker noise and DC offset are outside of the signal bandwidth.
A quadrature signal path is now needed for suppression of the image channels. Still, this
topology remains efficient in terms of power consumption.
Image rejection: Gain and phase mismatch in the quadrature paths leads to leakage
of signals around the image frequency into the wanted channel. Using zero-IF the signal
at the image frequency is the mirror of the wanted channel, and therefore the requirements
for image rejection are relatively relaxed in that case. However, when a low IF is used,
the signal at the image frequency of the wanted channel is its adjacent channel. Since
the adjacent channel may be significantly stronger than the wanted channel, the image
rejection requirements are much more severe for a low-IF than for a zero- IF topology
[104]. This implies that accurate gain matching and a precise 90 degrees phase difference
between the I and the Q paths is required. The image rejection (IR) is defined as:
sensitivity to the wanted channel
IR = (8.5)
sensitivity to the image channel
and can approximately be expressed as:
4
IR = 10 log (8.6)
A 2
A + (φ)2
where A/A represents the relative gain mismatch between I and Q, and φ is the phase
error in radians [105]. The FM/AM application requires about 80dB of image rejection.
The channel filter at IF (fig. 8.2) attenuates the adjacent channel (in FM mode) by about
40dB, so another 40dB of rejection must be provided by the matching of the I and the Q
path. Assuming equal contributions from gain and phase error, a gain error below 1.4%
is required, and the phase error must be less than 0.8degrees.
Integrated mixer and clocking scheme: A passive mixer can easily be integrated in
the input stage of a continuous-time ADC. In [38], it is shown that this technique
enables a high- resolution, high-linearity IF-ADC with low power consumption. Fig 8.12
shows how this technique is applied to the ADC for the FM/AM application. The input
signal around the 10.7MHz IF is applied as a differential voltage. It is converted to cur-
rents by input resistors Rin1 to Rin4 . Passive mixing is implemented by the switches at
the inputs of the ADC, which are low-ohmic (virtual ground) in this case. The on-
resistance of the switches is much smaller than the input resistance. Thus, high linearity
is achieved because the resistance of the switches is only weakly modulated. The power
consumption of the passive mixer is negligible.
The low-IF topology requires quadrature mixing and two parallel ADCs for the I
and the Q path. The combined output of both ADCs represents a complex signal that is
8.3. ADC DESIGN 127
Figure 8.12: Integration of a passive mixer into the FM/AM ADC and clocking
scheme
decimated in the digital, complex domain. In addition, the digital filter also performs the
final frequency shift to DC [106].
Fig. 8.12 also shows the clocking scheme for the ADC. The master clock at a fre-
quency of 42.14MHz is generated by an on-chip crystal oscillator. From this master clock,
the 21.07MHz sampling frequency for the ADC and the 10.535MHz frequency for
the mixer are derived. The latter frequency is provided twice, with a 90 degrees phase dif-
ference between the clock for the I and for the Q path. After the mixer, at the input of the
ADC, the signal is centered on the difference frequency of 165kHz. In the ADC, the
signal is 32 times over-sampled. At the output of the decimation filter a Nyquist sample
rate of 658kHz results.
Architecture
The ADC is based on the topology of fig. 3.4 and discussed in [7] and in section 3.2.2.
Here, the modulator is 32 times over-sampled and uses return-to-zero pulses in the feed
back in order to minimize inter-symbol interference. A 5th -order loop filter with complex
conjugated poles (i.e. local feedback loops with coefficients b3 and b5 ) is implemented
(see fig. 8.13). These poles provide additional filter gain within the signal bandwidth and
appear as notches in the shaped quantization noise spectrum. One of the notches is located
at the 165kHz IF frequency, the other one at the edge of the signal band. The loop filter
is implemented by means of simple transconductor-C integrators with feed forward coef-
ficients ai . These feed forward coefficients provide first-order roll-off at unity open-loop
gain for stability reasons. Large signal stability is guaranteed by clipping the integrator
Figure 8.13: 5th -order ADC
outputs, starting at the fifth integrator (graceful degradation). The resonators introduc-
ing the complex poles are implemented using local feedback transconductors gm,b1 and
gm,b2 . The resulting shaped quantization noise spectrum, assuming ideal circuit elements,
is shown in fig. 8.14.
In practice, various non-idealities affect the output spectrum. These are due to a.o.
the finite quality factor of the resonators, component spread and additional noise sources.
Quality factor of resonators: Transconductors gm2 , gm3 and b3 in fig. 8.13 constitute
the first resonator (located at the edge of the signal bandwidth). Taking into account the
output impedance of these transconductors, the resonance appears at:
1 + b3 R2 gm3 R3
ωr = (8.7)
R2 C2 R3 C3
where R2 and R3 represent the parasitic impedance that is in parallel with the integration
capacitors C2 and C3 , respectively. In case of a high-impedance node, eq. 8.7 is simplified
to:
b3 gm3
ωr ≈ (8.8)
C2 C3
In practice, the presence of R2 and R3 causes a shift of the resonance frequency to a lower
value.
Likewise, these impedances reduce the quality factor Q of the resonance to:
ωr
Q= (8.9)
1
R2 C2 + 1
R3 C3
By consequence, the notches in the noise shaping degrade.

8.3. ADC DESIGN 129
Figure 8.14: Simulated output spectrum with quantization noise, circuit noise
and jitter induced noise
Component spread: Component spread due to a variation in processing, tempera-

ture, supply voltage, etc., causes the noise shaping curve of fig. 8.14 to shift to the left
or the right. Table 8.3 lists the expected variation on the degeneration resistance and the
integration capacitance determining the time-constants of the loop filter for the 0.25µm-
CMOS technology the ADC is implemented in. From this, a 4σ -deviation of less than
25% is expected. Since the FM channel bandwidth typically equals 200kHz and the nom-
inal conversion bandwidth of the ADC is ∼ 330k H z, a worst-case negative spread of the
noise shaping in combination with a high operation temperature, can be accommodated.
Additional noise sources: In fig. 8.14, additional, white, noise sources are added to
the ideal quantization noise spectrum. The dominant noise source is due to the thermal
noise of the input and the feedback DAC resistors and the input transistors. (Each input
resistance equals 62.5k, each DAC resistance equals 100k.) As discussed in chapter 4,
this noise source can only be reduced at the expense of an increase in quiescent current
and therefore it is left to be dominant. The corner frequency of the flicker noise is outside
the bandwidth of interest.
Another limiting factor may be clock jitter. The approximation in [7] and simulations
show that the rms-value of the clock jitter must be below 7ps in order not to contribute to
the overall DR. This accuracy can be achieved with an on-chip crystal oscillator.
Circuits
The input stage of the ADC in fig. 8.13 is implemented as an active RC-integrator
(fig. 8.15). It is based on a simple single-stage OTA, with wide input transistors operating
Table 8.3: Variation on component values in the 0.25µm-CMOS technology
Processing spread Temperature [5; 125] deg C
P + -poly gate ox. P + -poly gate ox.

resistance capacitance resistance capacitance
4σ = 10% 4σ = 6% [−0.2%; +10%] < 1%
in weak inversion with gm = 4.3mA/V. It provides the virtual ground node where the
mixer can be conveniently implemented using N-MOS switches and where the DAC feed-
back current is subtracted from the input signal. Poly-silicon resistors and gate-oxide ca-
pacitors are used. Nulling resistors in series with the integration capacitors compensate
for the right-half plane pole that is introduced by this structure.
The P-MOS tail current source is degenerated with transistors in the triode region.
These perform the output common-mode control since they match with a replica circuit
that is biased at the target voltage.
The current consumption of the input stage, which is determined mainly by the noise
and the distortion requirements (see section 4.4), is 500µ A. Notice, P-MOS devices are
used in order to prevent back-gate modulation and to provide isolation from the low-
resistivity substrate.
The resonator stages consist of two gm C-integrators and a feedback transconductor.
The implementation is shown in fig. 8.16. All, use a degenerated, differential pair input.
The current consumption of a complete resonator is 240µA. The feed forward coefficients
are implemented as degenerated differential pairs as well. Their outputs are all fed to the
current input of the quantizer. It consists of a cross-coupled latch [7].
Decimation filter
The block diagram of the integrated, complex decimation filter is shown in fig. 8.17. The
21.07MHz 1-bit output of the I and the Q ADC is first filtered by a CIC filter and
decimated by a factor of 16. Then, the signal band of interest, around 165kHz, is shifted
to DC by a complex mixer (CORDIC algorithm, [106]) and, simultaneously, decimated
by a factor of 2. Next, a final decimation by 2 is performed to an output sample rate of
330kHz and a word-length of 16 bits.
Finally, it should be noted that careful layout of the DAC, the mixer and the input
stage is indispensable. Crosstalk from the large amount of quantization noise around half
8.3. ADC DESIGN 131
Figure 8.15: Input stage of ADC
the sampling frequency towards the IF input must be minimized. Layout symmetry is
essential to reduce gain and phase mismatch between the quadrature paths for best image
rejection.
8.3.3 Evaluation
A prototype IC was originally fabricated in a 0.35µm digital CMOS technology. It in-
cludes an I and a Q ADC with IF mixers, the complex decimation filters and the
digital shift to DC. Furthermore, a bandgap reference and a crystal oscillator are inte-
grated.
The IC is successfully scaled to a 0.25µm-CMOS technology. This resulted in
lower power consumption and a smaller chip area of 1.35mm 2 . The digital filters mea-
sure 0.75mm 2 and the analog part, including the bandgap is 0.55mm 2 . The oscillator is
0.04mm 2 . The ADCs and the decimation filter are indicated on the micrograph of fig. 8.9.
Measurement set-up
The measurement results of the next silicon in the 0.25µm-CMOS technology are re-
ported below. Both the 16-bit outputs of the digital block, at baseband, and the bitstream
outputs of the I and the Q ADC, at the low IF of 165kHz, are available for evaluation.
Differential, current-mode output buffers are used for the bitstream outputs in order to
minimize crosstalk. Below, though, the output spectra at baseband are shown.
The set-up is similar to that for the evaluation of the VGA in fig. 8.8. Of course,
the VGA is left out of the set-up. A HP8640B signal generator is used for low-noise
single-tone measurements and the IFR2026 generator is used for multi-tone measure-
ments requiring high linearity.
Figure 8.16: Resonator implementation
Figure 8.17: Block diagram of complex decimation filter
Measurement results
Fig 8.18 shows a complex output spectrum where the wanted signal is near the 165kHz
low-IF frequency. It is measured at the bitstream output of the ADCs. The maximum,
differential input signal of the ADC is 1.4V pp (for each ADC). Measured DR is 97dB in
9kHz (AM) and 82dB in 200kHz (FM) bandwidths. As expected, thermal noise domi-
nates in the bandwidth of interest.
Especially in AM mode, the peak-SNR is important, since a rise of the noise floor
caused by a large neighboring channel, decreases the sensitivity for the wanted channel.
Measured peak-SNR is 94dB in 9kHz at the maximum input signal of 1.4V pp . The 3dB
increase of the in-band noise at maximum input is probably caused by distortion of the
high-frequency quantization noise in the loop filter of the ADC.
Intermodulation distortion is especially important in AM mode, since the input signal
contains multiple AM channels. Distortion of strong unwanted channels may deteriorate
8.3. ADC DESIGN 133
Figure 8.18: Measured complex output spectrum
reception of the wanted channel. IM3 was measured by applying a full-scale, IF input
signal consisting of two 0.7V pp tones at 10.710 and 10.715MHz, so that the ADC input
signal is maximal. The resulting IM3 is −84d Bc (i.e. the intermodulation components are
84dB lower than the wanted signal) and decreases rapidly for small signals.
Non-linearity also leads to cross-modulation. This effect is well-known in AM re-
ceivers: when listening to a wanted channel that is not, or weakly, modulated, the modu-
lation of a strong unwanted channel may be heard, and thus disturbs the reception of the
wanted channel. Although this is an important test for AM-receivers, it is outside of the
present scope. For details, the reader is referred to [2]. The same is true for the evaluation
of image rejection. Finally, offset is measured. The test ICs show offset with σ = 0.7mV.
In the present application, this offset is outside the band of interest because of the low-IF
topology.
The total power consumption of the prototype IC is 36mW from a 2.5V supply volt-
age. Each ADC consumes 4mW and the complex digital decimation filter requires
11mW. The crystal oscillator consumes another 1.6mW. A major power consumption is
due to the current-mode, bitstream output buffers. These are for evaluation purposes only
and do not contribute during normal operation. Table 8.4 summarizes the main perfor-
mance parameters of the IF-ADC. Notice, the overall performance of the complex config-
uration is listed; i.e. the performance of the I and Q ADC operating together. Notice that
both FOMSINAD (eq. 4.4) and the modified, elaborate FOM of eq. 4.15 are listed. The first
FOM allows a first-order comparison with published ADCs. The second FOM is used for
comparison to the performance of the VGA in the previous section.
Table 8.4: Measured performance of complex ADC (I and Q together)
full-scale input 1.4V pp (differential)

input IF 10.7MHz
sample rate 21.07MHz
Nyquist sample rate 660kHz
AM (9kHz) FM (200kHz)
DR 97dB 82dB
peak-SNR 94dB 79dB
IM3 −84dBc
cross-modulation −91dBc
IR (20 samples) > 49dB
ADCs Dig. filters
Iquiescent 3.2mA @2.5V 4.4mA @2.5V
active area 0.55mm 2 0.75mm 2
FOMSINAD (eq. 4.4) 2.5 × 10−16 J
FOM of eq. 4.15 2 × 10−18 J
Discussion
The measurement results are in accordance with simulations and the design meets the tar-
get specifications. Hence, the presented ADC enables the radio architecture of fig 8.2. An
FM/AM radio with increased flexibility, higher quality and more features results. More-
over, an AM channel-select filter can be saved because of the multi-channel conversion in
this mode. This represents a significant cost reduction.
The successful realization of the IF-ADC and the presented results are the basis for
further development in this area. The design is further scaled to 0.18µm-CMOS and
adapted to encompass the IBOC digital radio standard. This is described in [107].
Benchmark
Table A.1 in appendix A gives an overview of published ADCs. Benchmarking the
AM performance is difficult because only few ADCs target a similar bandwidth. On the
contrary, many ADCs with a bandwidth of 200kHz have been reported and these are used
to benchmark the ADC in FM mode. From this comparison, it is clear that the presented
ADC combines a high DR and excellent linearity with very low power consumption. This
translates into the best FOM reported for a 200kHz-bandwidth ADC in CMOS technology
(to the authors knowledge).
8.4. EVALUATION OF THE CHANNEL 135
Comparing to ADCs with a bandwidth of 270kHz, mostly targeting GSM, the design
reported in [84] in 0.18µm-CMOS excels: it achieves a better FOM. That design uses a
lower supply voltage and larger input swing of 3V pp but these effects have already been
taken into account in eq. 4.15. Hence, the better power/performance balance may stem
from the larger over-sample ratio (see discussion on the Shannon theorem and power in
analog circuits on page 30).
8.4 Evaluation of the channel

The reported results are evaluated in relation to the discussion on power/performance re-
lations in chapter 4. Next, the presented channel for the highly-digitized FM/AM radio is
benchmarked with other published solutions.
8.4.1 Discussion
First, the power/performance relation achieved for the analog conditioning part (i.e. the
VGA) is compared to that of the ADC. Next, extrapolation of the overall power/perfor-
mance relation of the channel, in case of further digitization, is discussed.
Analog conditioning versus ADC
The presented VGA, in the highest gain setting, achieves a 10 times lower FOM (accord-
ing to eq. 4.15) than the ADC. This is mainly due to the fact that it achieves the same
DR for a smaller input signal (while its power consumption remains slightly lower). The
worse power/performance relation of the ADC may be due to various reasons:
• the loop filter consists of several stages requiring a minimum quiescent current for
robustness, matching, bandwidth, etc. The power consumed for these purposes,
however, hardly improves the power/performance relation in terms of DR and IM3 ;
• the ADC achieves higher performance, especially in terms of linearity. In general,

when the target performance of a circuit increases, more devices, other than the
input transconductor, start affecting the power/performance relation.
Still, this remark must be put into perspective: it is demonstrated in section 8.3.3 that this
ADC achieves state-of-the-art power/performance.
This comparison primarily indicates that a further digitization step, by reducing the
amount of VGA, will result in higher power consumption for the overall channel. The
ADC must then achieve even higher resolution for a smaller input signal. Because the
worse power/performance relation, the power increase in the ADC is higher than what is
gained from omitting the VGA .
Table 8.5: Benchmark of the channel in FM mode
analog filter[108] + PLL [103] band-pass this work

low-resolution ADC [102]
DR 61dB 85dB 79dB 82dB

IM3 a < −60dBc < −87dBc < −67dBc < −84dBc
P 12mW+P(ADC) 35mW 77mW+P(dig.) 19mW
a at full-scale input
ADC versus digital conditioning

The power consumption of the digital part of the channel is just slightly higher than that of
the ADC. The situation corresponds to the graph in fig. 4.4. The graph indicates that,
in case a further increase of resolution were required (for instance for further digitization),
the power consumption of the ADC would become dominant while that of the digital
part increases much less.
From the discussion in this section, it becomes clear that further, straightforward
digitization of the FM/AM channel will result in a power increase for the overall channel.
As depicted in the graph of fig. 4.4 the power consumption in the ADC becomes dominant
and increases dramatically.
8.4.2 Benchmark
This benchmark focuses on the performance in FM mode because more publications are
available than for AM operation. In fact, to the author’s knowledge, the presented work
is the first solution reporting multi-channel A/D conversion in AM radio receivers.
Benchmark with highly-analog channels

[108], [109], [110] and [111] report integrated, band-pass filters around the 10.7MHz
IF. A high-quality integrated filter reduces the required external filtering and relaxes the
consequent A/D conversion. [108] achieves the lowest power consumption and, therefore,
it is listed in table 8.5. Power-wise (anticipating a moderate power consumption in the
“low-resolution” ADC) it may enable a more efficient channel than the presented solution
with the IF ADC. Performance-wise there is a significant gap, though: this filter
targets application in portable radio and achieves less DR and linearity.
[103] presents a PLL performing analog demodulation and digitization of the
FM signal. It achieves excellent performance but does not allow future development to
encompass AM radio as well.
It is concluded that the power/performance balance of the continuous-time ADC

is competitive or better than that of an analog channel with a cascade of dedicated, analog
conditioning blocks.
Benchmark with highly-digitized channels

[90] reports the use of a wide-band ADC for digitizing the FM/AM channel. The ADC
achieves 10 bits of resolution in 11 MHz bandwidth. Power consumption is not men-
tioned.
[96], [97], [98], [99], [100], [101] and [102] present band-pass converters digi-
tizing the signal at the IF. Among these, [102] achieves the best overall performance. Still,
its power consumption is significantly higher than that of the other solutions in table 8.5.
Moreover, the power consumption in the decimation filter is not yet included. In addition,
linearity is difficult at these frequencies.
In FM mode, the presented channel with passive mixing to a low IF and consequent
low-pass, continuous-time A/D conversion achieves a better overall power/perfor-
mance relation than the other solutions in table 8.5. This ADC especially excels in terms
of linearity. Therefore, it is the only design allowing the multi-channel conversion of
fig. 8.2 in AM mode. All other solutions in table 8.5 have too much distortion and would
not be suitable.
8.5 Conclusions
A dual-mode conditioning channel for an FM/AM radio receiver has been presented. In
FM mode, the channel remains highly analog. In AM mode, multi-channel A/D conver-
sion enables highly-digitized signal conditioning with digital channel selection. Two key
sub-circuits have been presented.
A CMOS VGA operating at a 10.7MHz IF frequency, achieves −67dBc IM 3 -distortion

at 1.4V pp output. This is competitive with state-of-the-art designs but does not meet the
requirements for the present application in AM mode.
A continuous-time ADC achieves a FOMSINAD of 2.5 × 10−16 J (from eq. 4.15)
and features a very low IM3 -distortion of −84dBc for 1.4V pp input signal. It consumes
only 8mW.
The evaluation of these blocks leads to the conclusions that are listed below.
The power/performance balance of the continuous-time ADC is competitive or

better than that of dedicated FM-solutions consisting of a cascade of analog conditioning
blocks.
In case of further digitization, the power consumption of the ADC becomes domi-
nant and causes a dramatic increase of the power consumption of the overall channel.
The multi-channel aspect, makes the operation in AM mode more challenging than
in FM mode, especially because of the very high linearity requirement.
The presented ADC enables the first solution reported for multi-channel A/D conver-
sion in AM radio receivers. As such, it demonstrates the feasibility of a highly-digitized
channel for narrow-band systems, even while the linearity challenge is very demanding.
These conclusions illustrate the theory of chapters 4 and 5.

Chapter 9
Conditioning ADCs for

Bluetooth
This chapter presents the application of “conditioning ADCs” in a Bluetooth receiver.

Design examples in sections 9.2, 9.3 and 9.4 illustrate the theory of section 6.2 on the
feed forward ADC with signal conditioning in the decimation filter, of section 6.3 on the
conditioning feedback ADC with restricted filtering and of section 6.4 on the FFB-ADC,
respectively. Each of these designs enables a Bluetooth-compliant receiver without the
need for analog conditioning in front of the ADC. This is different from other published
Bluetooth receivers. It is shown that, integrating explicit signal conditioning into the
ADC (as done in the designs of section 9.3 and 9.4) promises a major improvement in
the overall power/performance ratio of the receiver. This is true both when comparing to
a “conventionally digitized” receiver as well as when comparing to a dedicated analog
solution.
9.1 System
Bluetooth [112] is a short-range wireless communication system targeting operation in
the 2.4GHz ISM band. In most countries, the channels occupy the band from 2.402 to
2.480GHz. The channel spacing is 1MHz and GFSK modulation is used. The latter im-
plies that all information is contained in the phase of the signal. The Bluetooth standard
targets a moderate data-rate (i.e. 1Mbps for the early standard) for use in portable appli-
cations as for example headsets, printers, etc. By consequence, cost and power drain are
the main challenges on the IC solution.
For cost reasons especially, digitization is welcomed since it enables mass-production
in an advanced CMOS technology. The key question, however, is whether a digitized so-
lution is competitive with an analog solution in terms of power consumption.
Another motivation for digitizing the Bluetooth receiver, results from the fact that ex-
tended standards like Bluetooth Medium-Rate and High-Rate are being defined in order to
provide higher data-rates. In these standards, the channel spacing remains unaltered and,
139
140 CONDITIONING ADCS FOR BLUETOOTH
Figure 9.1: Bluetooth receiver with analog (a) and with digital demodulation
(b)
therefore, a higher-order modulation scheme including amplitude modulation is required

to enable this increase. For example, the medium-rate standard targets 2Mbps and uses
π/4-DPSK modulation. This results in a non-constant envelope signal and the simple
analog demodulation schemes are not applicable anymore.
9.1.1 Conventional radio with analog demodulation

A conventional Bluetooth receiver with analog demodulation is depicted in fig. 9.1.a. In
this application, it is common practice to use a low-IF topology to circumvent the problem
of DC errors falling into the signal bandwidth. The baseband part of the receiver consists
of a cascade of filter and gain sections. The latter may be clipping to limit the signal to
a pre-defined level and can be used to indicate the received signal strength. The bits are
recovered by further analog demodulation.
Plenty of low-power, analog demodulation schemes are available for the “simple”
GFSK modulation that is used for Bluetooth. They often originate from FM demodulators
and allow an easy implementation, resulting in a small area and low cost. For example, a
frequency discriminator or a phase-locked loop detector can be used [113]. Alternatively,
many zero-crossing detectors have been reported [114], [115], [116] a.o.
9.1. SYSTEM 141
Even though these analog demodulators are robust and simple they require a lot of
preceding analog conditioning. More important, the analog receiver architecture is not
very future-proof: adaptation to include amplitude demodulation -required in the extended
standards- is not straightforward.
9.1.2 Radio with digital demodulation and analog

signal-conditioning
Alternatively, digital demodulation can be implemented [106], [117] as in fig. 9.1.b. Note,
the conditioning channel is linear; i.e. the amplitude information is preserved and can
be used to improve the GFSK demodulation or to encompass extended standards (with
amplitude modulation).
Various Bluetooth receivers with digital demodulation have been reported in a.o.
[118], [119], [120] and [121]. They differ in the degree of digitization of the signal-condi-
tioning. For example, [118] uses a 7th -order filter inter-leaved with 84dB of VGA in steps
of 1dB, followed by a very simple ADC. Both the filter and the VGA need calibration and
tuning to handle spread and drift on the analog parameters.
9.1.3 Radio with digital demodulation, without analog

signal-conditioning
A much higher degree of digitization is reported in [121]. The work presented in the re-
mainder of this chapter is part of the overall receiver design described in this reference.
The referenced receiver does not need any baseband analog channel filters or VGA. In-
stead, a high-resolution ADC (see section 6.2) with signal-conditioning in the decimation
filter is used. This first solution is discussed in section 9.2. On top of this, in the second
solution (section 9.3), some restricted signal-conditioning is integrated into the ADC. In
the third solution (section 9.4) the signal-conditioning in the ADC is unrestricted. The
other sections in this chapter discuss implementations with more signal conditioning in-
tegrated into the ADC.
The single-chip receiver [121] of which the current conditioning channel is a part of,
is further detailed in fig. 9.2. Two LNAs are used in parallel to provide sufficient isola-
tion between the I and the Q channels at the mixer input. The LNAs are simple V-to-I
converters. The RF output current is down-converted to a low IF frequency of 500kHz
by passive mixers, directly driven by a quadrature VCO in a PLL. Hence, the wanted
channel extends from 0 to 1MHz at the ADC input. A ADC is used in combination
with signal-conditioning in the decimation filter, as described in section 6.2; i.e. the ADC
provides high resolution within the bandwidth of the wanted channel while the interferers
are present in the noise-shaping part of the output spectrum. The digital block performs
channel and decimation filtering as well as demodulation.
The above discussion was generic. From now on, we focus on the radio as intro-
duced in section 9.1.3. More specifically, the signal conditioning at the low-IF frequency
is considered. Obviously, this receiver achieves a higher degree of digitization than that of
fig. 9.1.b. Because of the absence of analog channel filters and VGA stages, the receiver
Figure 9.2: Bluetooth receiver with digital demodulation, without preceeding

analog filters or VGA
hardly needs tuning and is much more flexible. On the other hand, though, this absence
of analog conditioning has major implications on the ADC specifications:
• high DR: The ADC must handle the full signal range. The wanted signal can be
very small because of the lack of preceding gain. Even then, the ADC must provide
sufficient SNR to enable digital demodulation.
• interferer immunity: Interferers are present at the ADC input but they should not
affect the resolution in the wanted channel. Therefore, the ADC must provide alias-
suppression, it must be highly linear and interferers should not cause overload or
spurious responses.
• bandwidth requirements: The interferer bandwidth can be 78MHz (corresponding

to the bandwidth of operation for Bluetooth) while the sample rate of the ADC is
only 64MHz. Hence, the bandwidth requirements are primarily set by the interfer-
ers.
• termination of RF: The ADC is directly connected to the RF front-end. It is in

series with the passive mixer and acts as a load to the LNA. Since the LNA has
a current-mode output the ADC must provide a low-impedance termination, i.e.
its input impedance must be smaller than 400 over the entire 78MHz Bluetooth
band.
9.2 Feed forward ADC

A feed forward ADC in combination with signal-conditioning in the decimation filter (see
section 6.2) has been implemented for this Bluetooth receiver. As a short-hand notation, it
is referred to as the “feed forward ADC”. Further on, it is used as a reference to compare
other conditioning ADCs to. At the output, a SNR of only 18dB is required for
9.2. FEED FORWARD ADC 143
Figure 9.3: Quadrature conversion using two low-pass ADCs (a) or using a
single complex ADC (b)
digital GFSK demodulation. Since the amplitude of the wanted signal may vary by 50dB
and taking into account 5dB of margin, the ADC must then provide 73dB of DR in a
bandwidth of 1MHz around the 500kHz low-IF frequency. The linearity requirement is
set by the interferers. From the Bluetooth standard it can be derived that IM3 distortion up
to −60dBc can be allowed. Image rejection can be very moderate because the Bluetooth
standard allows for an exception here: it is assumed that the image channel is not occupied
by an interferer. The major challenge on the Bluetooth ADC is in achieving a high DR at
low power consumption while being robust to the presence of interferers.
9.2.1 Design
Next to the above performance targets, system level choices in the receiver design set an
additional boundary condition on the ADC architecture. For example, the system clock
frequency is chosen as low as possible in order to avoid an additional PLL. Instead, it can
be generated by an on-chip oscillator tuned at the third over-tone frequency of a crystal,
generating 64MHz. Unfortunately, this clock frequency limits the over-sampling of the
ADC. In order to meet the DR requirement either multi-bit quantization or very aggressive
noise-shaping (i.e. a high-order loop filter) is needed. For reasons of linearity and power,
the latter option is preferred (see discussion in section 3.2.1) and a 5th -order loop filter is
used.
ADC architecture
The signal bandwidth from 0 to 1MHz can in principle be converted into the digital do-
main using two identical low-pass ADCs processing the I and Q channels separately
(see fig. 9.3.a), just as for the FM/AM receiver in the previous chapter. The noise transfer
Figure 9.4: Comparison between real (fig. 9.3.a) and complex noise-shaping
(fig. 9.3.b) in graph (a) and effect of a small mismatch between the
I and the Q path in graph (b)
Figure 9.5: Effect on noise-shaping when putting one notch in the image band
function of the ADCs is symmetrical with respect to DC, in this case, and the same res-
olution is available from −1MHz to 0Hz as from 0Hz to 1MHz, which is not necessary.
Hence, the idea to use a single complex ADC with quadrature inputs and outputs
and with a poly-phase loop filter to achieve band-pass noise-shaping (see fig. 9.3.b). The
noise-shaping of both topologies is compared in fig. 9.4.a. Complex noise-shaping is
more efficient: because the quantization noise can be higher in the image band, it can
be deeper suppressed in the wanted channel. In other words, the effective over-sampling
has doubled. In view of the rather low sample frequency, complex noise-shaping seems
particularly attractive.
One drawback of high-order, complex noise-shaping is in the risk of “leakage of
quantization noise”. The ideal, complex noise-shaping of fig. 9.4.b is only achieved in
case of perfect matching between the STF of the I and the Q path of the ADC. In
practice, mismatch causes quantization noise, present in the image channel, to leak into
Table 9.1: Simulated SQNR of various 5th -order architectures (bandwidth

is 1MHz, sample rate is 64MHz)
real complex, matching is: complex, 1 notch

(2 ADCs) perfect -45dB in image channel
SQNR 77dB 103dB 82dB 90dB
the wanted channel. Because of the very high noise at negative frequencies, this has a
disastrous effect on the SNR in the wanted channel. Fig. 9.4.b illustrates this effect in case
of a small gain (or phase) mismatch of only -45dB between the STF of the I and the Q
path.
A solution to the leakage problem was introduced in [97]: instead of putting all filter
notches in the wanted channel, one notch is left in the image channel (fig. 9.5). Obviously,
the quantization noise in the wanted channel increases slightly but now, mismatch up to
-20dB can be tolerated. Hence, the overall robustness of the architecture has improved
considerably.
Table 9.1 compares the signal-to-quantization-noise ratio (SQNR) that is achieved by
the various architectures. The complex noise shaping with one notch in the image
band improves the SQNR by 13dB compared to the case when using two ADCs with real
noise-shaping. This improvement is essential in order to achieve the 73dB target in the
presence of other noise sources like circuit noise and clock jitter.
Fig. 9.6 shows the block diagram of the ADC with the 5th -order, complex loop filter.
Starting from two low-pass filters with 90 degrees phase-shifted input signals, a band-pass
filter is constructed by applying cross-coupling paths between the I and the Q filter at the
output of the fourth and the fifth integrator. These cross-coupling paths realize complex
filter coefficients and the filter response becomes asymmetric around DC. The notches
appear at the following frequency:
gm,e4 gm,e5
2πC4 ≈ 750k H z and 2πC5 ≈ 300k H z
The local feedback from the output of the third integrator to the second integrator in-
troduces a pair of complex conjugated poles. This results in a notch at the edge of the
signal channel and one at the edge of the image channel (i.e. at +1MHz and at -1MHz)
and provides additional suppression of the quantization noise here. The resonance fre-
quency can be calculated from eq. 8.7. Fig. 9.7 shows the simulated magnitude of the
loop filter response. The finite output impedance of the various integrators limits the peak
of the notches (i.e. reduces the quality factor of the resonance) and shifts the resonance
frequency to a marginally lower value than what has been calculated.
Figure 9.6: 5th -order ADC with complex loop filter
ADC circuits
The design of the various OTAs is similar to the topologies of fig. 8.15 and fig. 8.16.
These have been ported to a 0.18µm-CMOS technology. Despite of the multiple stacked
transistors this topology is successfully implemented within the 1.8V supply rail. Be-
cause it requires only a minimum number of current branches it remains preferred over a
folded cascode OTA or any two-stage topology. Fig. 9.8 shows the nominal bias voltage
of various nodes and demonstrates the available headroom.
Contrary to the implementation in section 8.3 in the 0.25µm-technology, here, N-
MOS devices are used for the input transistors. In the 0.25µm-technology, with a low-
impedance substrate, P-MOS devices were chosen for isolation reasons. The 0.18µm-
technology uses a high-impedance substrate (i.e. the resistivity equals 10cm) and there-
fore this argument becomes less important. On the other hand, at the same quiescent
current, the N-MOS transistor yields a three times higher gm than the P-MOS transistor
Figure 9.7: Magnitude of loop filter transfer
Figure 9.8: DC biasing of first integrator

Table 9.2: Variation on component values in the 0.18µm-CMOS technology
Processing spread Temperature [5; 125] deg C
N + -poly gate ox. N + -poly gate ox.

resistance capacitance resistance capacitance
4σ = 17% 4σ = 10% [−1%; +3%] < 1%
and is preferred for this reason. Since the N-MOS transistor is referenced to ground (via
its bulk connection), all other signal-conducting devices are referenced to ground as well
such that any substrate bounce remains common mode.
The time-constants of the loop filter depend on the product of a N + -poly resistance
and gate-oxide capacitance. Table 9.2 lists the expected variation on these parameters.
The overall, worst-case spread on the RC-product is less than 25%. In the FM/AM ADC,
the spread was dealt with by providing a bandwidth margin. Here, the spread is com-
pensated for by adding or removing integrator capacitors in discrete steps. Therefore, the
filter bandwidth is made programmable in four steps allowing a correction of −12.5%
up to +25%. Note that this interval is asymmetric. Especially the case where the filter
bandwidth is smaller than intended needs correction since, then, the quantization noise
shifts into the wanted channel. The calibration algorithm is out of the scope of this book.
Fig. 9.2 implies that the ADC has a current-mode input. The input is directly driven
by the output current of the LNA that is down-converted by the passive mixer (similar to
the mixer in the FM/AM receiver and discussed on page 126). The ADC must provide
a low-impedance termination of the RF front-end to guarantee a linear operation of the
LNA and the mixer. It is calculated that the differential input impedance of the ADC must
be below 400 over the entire bandwidth of the input signal, including interferers (i.e.
over 78MHz). The differential input impedance of the ADC is calculated:
2
Z in ( jω) ≈ F( jω)
(9.1)
Rfb + gm1 + jωC gate
Most parameters in this formula have been defined in fig. 9.6, others are:
F( jω) = transfer function from v,Q to v D AC,Q in fig. 9.6
(or from v,I to v D AC,I )
C gate = gate capacitance of first OTA
At low frequencies, the input impedance approximates 2Rfb /F(jω) and is very small due
to the high loop gain. Beyond the unity-gain bandwidth of the loop the input impedance
Figure 9.9: Differential input impedance of ADC as a function of frequency

(peaking is due to numerical inaccuracy)
of the ADC equals that of the first stage, i.e. 2/gm1 . For very high frequencies, Z in is
limited by the gate capacitance of this stage. The magnitude of Z in is plotted in fig. 9.9
for positive frequencies. The differential input impedance remains below 400 for all
frequencies.
In this particular receiver architecture, the requirement for a low-impedance termi-
nation sets the quiescent current of the first stage and dominates the power consumption
of the overall receiver. This quiescent current is so high that -contrary to the generic
assumption in the power/performance relation of eq. 4.13- here, the noise and distortion
of the first stage is less important than that of the feedback resistor R f b and of gm2 . This
would not be the case for an optimized, stand-alone ADC.
The circuit noise is significantly larger than the quantization noise and limits
the achievable DR to 79dB. On top of that, jitter on the DAC clock translates into an
additional noise contribution. The short-term jitter of the 64MHz-oscillator amounts
7 psr ms (from measurements). Assuming white noise and based on [7] it can be calculated
that, due to the jitter, the overall DR of the ADC further degrades to 76dB.
Digital conditioning and demodulation
Fig. 9.10 shows a functional diagram of the digital conditioning blocks and the demod-
ulation. A CIC filter (see eq. 4.17 with k = 6) performs decimation filtering and down-
sampling by a factor 8. The output sample rate equals 8MHz. A rotating CORDIC-block
[106] shifts the IF frequency of the outputs of the CIC filters from 500kHz to DC. The
consequent matched filter provides optimal suppression of white noise. In addition, it also
Figure 9.10: Functional diagram of digital part of the receiver of fig. 9.2
suppresses the remaining quantization noise from the ADC and therefore can be con-
sidered as a decimation filter as well. The remaining digital blocks perform demodulation
and compensation of a potential frequency offset [121]. These functions are beyond the
scope of this book.
9.2.2 Evaluation
The ADC has been embedded in a single-chip Bluetooth receiver (see fig. 9.2), designed
in a 0.18µm digital CMOS technology. Fig. 9.11 shows the die micrograph with the
various receiver blocks. The evaluation of the overall receiver is documented in [121].
Highlights are the sensitivity of -71dBm at a power consumption of less than 32mW, the
area of 3.5mm 2 and the IIP3 of -11dBm achieved without any analog filtering or VGA.
Here, the focus is on the performance of the ADC, though. The evaluation of a
stand-alone version is presented. The test IC includes the complex ADC, a bandgap
reference, the crystal oscillator and low-swing buffers to output the bitstreams. Resistors
of 10k are added in series with the input terminals to facilitate evaluation. This value is
comparable to the output impedance of the RF front-end. While in the actual receiver the
maximum differential input signal to the ADC (i.e. the output of the mixer) is 100µA pp ,
it equals 1V pp in this set-up.
Set-up
An Arbitrary Waveform Generator (Tektronix AWG420 with 16-b word-length and a

sample rate up to 200MSps) is used to generate the differential, quadrature input sig-
nals with sufficient resolution. For some tests, external calibration of the generator is
required to reduce harmonic and intermodulation distortion of the generator. For tests
with high-frequency interferers, high-pass filters (Mini-Circuits BHP-25) with stop-band
until 25MHz are used to suppress the low-frequency spurious from the generator. The
bitstream outputs are grabbed with a VXI-analyzer and further processed in LabView.
Figure 9.11: Micrograph of the single-chip Bluetooth receiver
Conventional performance parameters

First, some conventional ADC performance parameters such as SNR, linearity and image
rejection are discussed.
Conventional performance parameters: Fig. 9.12.a shows the SNR as a function of

the digital output: a peak-SNR of 75.5dB and a DR of 76dB are achieved. Fig. 9.12.b
shows the result of a two-tone test: IM3 -distortion is below −82dBc relative to the input
tones, each at half of the full-scale input. IR is typically better than 50dB and is mostly
dominated by a gain mismatch. The worst-case value, measured over 17 samples is 47dB.
This is far better than required. Further statistics on the measured image rejection are
quoted in [32].
The target specifications as defined in the introduction, are all well met. Table 9.3
gives an overview of the conventional ADC performance parameters and evaluates the
FOMSINAD of eq. 4.4. The latter is discussed in section 9.5.
Interferer immunity
Because of the lack of preceding filters in this particular receiver topology, the ADC im-
munity with respect to interferers must be evaluated next.
Aliasing limit on interferer immunity: In a first test, an interferer is applied near the
sample frequency of the ADC in order to evaluate the aliasing suppression. The input level
of the interferer complies with the blocker level defined in the Bluetooth standard [112].
Figure 9.12: SNR as a function of digital output (a) IM3 -test with f1 = 530kHz
and f2 = 730kHz (b)
Table 9.3: Measured performance of the complex ADC (conventional parame-

ters)
full-scale input 100µA pp (differential)

bandwidth 0-1MHz
sample rate 64MHz
DR 76dB
peak-SNR 75.5dB
IM3 < −82dBc
I R (17 samples) > 47dB
Iquiescent 2.5mA @1.8V
FOMSINAD eq. 4.4 1 × 10−16 J
The interferer frequency is chosen slightly higher than the clock frequency such that the
alias appears at a positive frequency in the bandwidth of the wanted signal (fig. 9.13). (In
case the interferer frequency was lower than the sample frequency, the alias would appear
at a negative frequency.) The alias component is below the allowed limit of −68dBFS de-
rived from the Bluetooth blocker tests1 . In general, for all measured samples, the aliasing
1 The test defines a blocker at -27dBm, corresponding to -10dB of the ADC full-scale output, while the
wanted signal is at -67dBm, corresponding to -50dB of the ADC full-scale output. Since 18dB of SINAD is
Figure 9.13: Aliasing test with an input signal at +370kHz offset from the sample
frequency at a signal level of −10dBFS
suppression exceeds 60dB. Still, this is lower than what is expected from eq. 6.3. This
discrepancy is probably due to the fact that -next to the aliasing- also parasitic mixing via
cross-coupling occurs. As such, the interferer is mixed back into the signal band.
Stable input range for interferers: A second test concerns the stable input range for
interferers. As discussed in section 6.1.3, this limit is frequency dependent, being in-
versely proportional to the STF of the ADC. The simulated limit as well as the measured
stable input range over frequency is depicted in fig. 9.14. The simulated limit is based on
an AC analysis of the implemented ADC where DAC and quantizer have been modeled
by a linearized gain. The simulated limit deviates from the measured curve, especially for
adjacent channels. This is due to numerical problems caused by a fast transition of the
amplitude and the phase of the loop gain. In addition, the simulated limit is based on a
linearized model assuming small-signal operation. This model becomes inaccurate near
full-scale and, therefore, should be considered as a first-order estimate only. Better cor-
respondence can be achieved if transient simulations are used to predict the stable input
range. The penalty of that approach is in a very long simulation time.
The measured graph is asymmetrical around DC. This is due to the complex nature of
the loop gain and the STF. At positive frequencies, the stable input range is large enough
to accommodate the interferers defined in the various Bluetooth tests and indicted as the
“Bluetooth mask”. At negative frequencies, the margin is tight, though. For this test, only
five samples have been evaluated because the frequency sweep needs to be performed
manually. (Because of the differential quadrature inputs, an automatic set-up would re-
required for demodulation, the alias needs to be below −68dBFS .

Figure 9.14: Stable input range over frequency
quire four programmable attenuators of sufficient performance. These were not available
at the time.) The variation on the measured stable input range is small. All measured
samples accommodate the Bluetooth mask as depicted in fig. 9.14.
Spurious responses due to interferers: Thirdly, an interferer around half the sample
frequency, i.e. 32MHz is applied. In terms of spurious responses, this is a worst-case input
since the modulator tends to make correlated patterns at this frequency. It is evaluated
that, as long as this interferer remains smaller than 0.1V (this corresponds to -17dB on
the digital scale), the spurious responses within the wanted channel do not affect the DR
of 76dB.
IM3 distortion of interferers: Finally, IM3 distortion is evaluated over frequency:

interferers are applied at frequency f 1 and f 2 , such that the intermodulation component
2f1 −f2 falls in the middle of the wanted channel, i.e around 500kHz. The interferers
f 1 and f 2 are applied at an input level equaling half the maximum stable input at those
frequencies. This input level may differ from that of a conventional IM3 -test for wanted
signals where two inputs at −9dBFS (adding up to −3dBFS ) are applied. In case interfer-
ers were applied at the latter input level, they would either cause compression (in case the
stable input range is lower than −3dBFS ) or the IM3 -components would be hidden in the
noise (in case the stable input range corresponding to this input frequency is much higher
than −3dBFS ).
It was mentioned in section 4.4 that the transconductors of the input stage are the
dominant source of distortion in a feed forward ADC. This is true for the wanted chan-
nels. Fig. 6.16 shows that, especially when applying adjacent interferers, the internal
signal swings can become large. For these frequencies, distortion at the output of the
first integrator may become important as well. This is evaluated next. Afterwards, IM3 -
Figure 9.15: IM3 test with adjacent interferers: f1 = 2MHz, f2 = 1.25MHz (a)
and with far-off interferers: f1 = 30.25MHz, f2 = 60MHz (b)
distortion of far-off interferers is measured.

Fig. 9.15.a shows the IM3 -distortion when applying adjacent interferers
f1 = 1.25MHz and f2 = 2MHz, each at −12dBFS . As such, the interferers add up to the
maximum stable input of −6dBFS for this frequency range. The distortion component
at 500kHz is below −92dBFS . Note the low-frequency spurious components at 250kHz,
750kHz, etc. These are due to the generator. The component generated at 500kHz has
been suppressed by calibration of the generator such that only the contribution of the ADC
appears in the output.
As mentioned, distortion is lower for far-off interferers [112]. This is shown in
fig. 9.15.b with inputs at f1 = 30.25MHz and f2 = 60MHz. In this set-up, high-pass filters
with the stop-band up to 25MHz suppress the low-frequency spurious generated by the
AWG. Again, the IM3 component falls at 500kHz. Both inputs are applied at 0dBFS and
add up to +6dBFS . (This is 9dB higher than in a conventional measurement because,
on one side, the stable input range is high enough to allow for this input level and, on
the other side, applying a signal at the conventional input level would require too many
samples to detect the IM3 component). Clearly, the distortion is low enough not to limit
the allowable interferer level for these frequencies. In fact, the allowable input for far-off
interferers is limited by the supply voltage. Also note the tones around 1.75MHz (and
-1.75MHz): these spurious components are caused by the large input near half the sample
frequency; i.e. these are spurious responses. They must be suppressed in the decimation
filter.
Fig. 9.16 summarizes the evaluation of the interferer immunity of the presented ref-
erence feed forward ADC. For reference purposes, it is repeated that the input amplitude
of a full-scale wanted signal equals 0.5V , corresponding to −3dB on the digital scale.
This graph corresponds to the simulations on the 3r d -order ADC discussed in chapter 6
(fig. 6.8). For adjacent channels, the allowable interferer level is limited by the stable
input range of the ADC. Around 32MHz, the interferer level must not exceed 0.1V (this
corresponds to -17dB on the digital scale) to maintain a DR of 76dB in the wanted chan-
Figure 9.16: Overall limit on the allowable interferer level over frequency
nel. Around 64MHz, interferers up to −19dBFS are allowed: in view of the aliasing or
parasitic mixing of −60dB this input level does not affect the DR of 76dB. For all other
frequencies, the allowable interferer level is set by the supply voltage. This limit corre-
sponds to +3dB on the digital scale; i.e. an interferer with an amplitude of 1V . Note,
distortion is not a limiting factor for the interferers applied to this ADC.
9.3 Conditioning feedback ADC

In this section, a ADC with feedback compensation is presented as an alternative
solution for the receiver baseband of fig. 9.2. The feedback topology is used in view of
its filtering STF and allows signal-conditioning integrated into the ADC (see section 6.3).
The feedback ADC targets the same performance as the feed forward ADC in the previous
section and has been implemented in the same 0.18µm CMOS technology. Moreover, the
same circuit topologies have been reused for the various building blocks.
The asset of the feedback ADC is in its improved interferer immunity. The presented
design illustrates the theory discussed in section 6.3. In particular, it clarifies the trade-off
between:
• an improved interferer immunity;
• a competitive power/performance ratio.
This trade-off is reflected in the discussion in section 9.3.1 and is is evaluated based on
measurement results in section 9.3.2.
9.3. CONDITIONING FEEDBACK ADC 157
9.3.1 Design
Since the feedback ADC targets the same DR as the feed forward reference ADC, es-
sentially, the same NTF needs to be realized. In view of interferer immunity, though, a
different -i.e. filtering- STF is pursued. It is shown that, in order to come to a power-
efficient implementation of the feedback ADC, some concessions to the ideal NTF and
STF must be made. Afterwards, some circuit details are presented.
ADC architecture
The feedback ADC can be designed starting from the same NTF as for the feed forward
reference ADC of section 9.2, i.e. the same function L 1 (see chapter 6) is realized but it
is mapped on a filter with feedback compensation. In section 6.3, it was shown that this
yields a filtering STF but it also results in unfavorable values for the unity-gain frequen-
cies of the integrators in the loop filter.
Example
Assuming the same boundary conditions as for the feed forward design (i.e. the same NTF,
the same maximum input signal and the same allowable swing at the integrators’ output)
a unity-gain frequency of 300kHz is required for the first integrator in the feedback topol-
ogy compared to 2.8MHz in the feed forward topology. This implies that in the feedback
topology the first integrator has attenuation inside of the signal bandwidth of 1MHz. Ob-
viously, this is very disadvantageous for the overall power/performance balance.
This problem is reduced by making the concessions listed next.
• A 4th order instead of a 5th order loop filter is implemented: the unity-gain fre-
quency of the first integrator can then increase to 600kHz. (Since less feedback
paths are present and thus less signal is injected in the loop the first integrator can
have higher gain.) The penalty of this choice is in a reduced SQNR of about 80dB
instead of 90dB for the 5th order design. In addition, the interferer immunity is
slightly affected since most of the limitations discussed in section 6.1.3 depend on
the loop filter order.
• A local feed forward path replaces a feedback path: this is illustrated for a simple
3r d order example in fig. 9.17.a and b. It can be calculated that both topologies are
equivalent if the local feed forward coefficient equals d2 /d1 and if the first and the
second integrator are interchanged. As such, the unity-gain frequency of the first
integrator becomes 1.9MHz. The penalty of this adaptation is in a further reduction
of the slope of the filtering STF by one order and in some overshoot in the adjacent
channel. In addition, the feed forward path reduces the aliasing suppression by one
order and affects the immunity to spurious responses.
Both measures compromise on the interferer immunity but improve the power/perfor-
mance balance by allowing more gain in the first integrator of the loop filter. (Note that
the gain of the first integrator is still lower than in the feed forward implementation.) This
Figure 9.17: Conventional feedback topology (a) and modification with a local
feed forward path (b)
Figure 9.18: Simulated STF of a 5th -order feed forward, 5th -order feedback and
a 4th -order modified feedback ADC
compromise is evaluated in section 9.3.2 when presenting the measured performance. For
now, fig. 9.18 compares the STF of the presented topology to that of the 5th -order feed
forward ADC of the previous section and that of a 5th -order feedback ADC. These are
simulated curves based on a linearized model of the various topologies. The filtering
behavior of the present topology is not as good as that of a 5th -order implementation with
“conventional feedback”. It is still an improvement on that of the feed forward topology.
Fig. 9.19 depicts a block diagram of the implemented, “modified” feedback ADC. As
in the previous ADCs, some local resonators have been implemented to create additional
loop gain at particular frequencies. The pair of complex conjugated poles hence creates
a notch at +1MHz and at -1MHz (for suppression of quantization noise in the image
Figure 9.19: 4th -order ADC with modified feedback topology
channel). The cross-couplings between the I and the Q path of the loop filter create
notches at 200kHz and at 600kHz. The variability of Rin is discussed next. The variability
of Rin is discussed next.
Integration of programmable gain

Because of the order of its loop filter, this ADC achieves a lower SQNR than the feed for-
ward ADC in the previous section. In addition, the thermal noise of this ADC is somewhat
higher in order to save on power consumption. By consequence, the design of fig 9.19 has
a lower DR, and a lower peak-SNR. However, this topology excels in interferer immunity.
Hence, the option to extend its DR by integrating some programmable gain: the input
resistance is switched from 100k, to 10k and to 1k depending on the amplitude of
the wanted signal2 . This results in a dynamic adaptation of the input range to the strength
of the wanted signal and in a larger input-referred DR.
2 Information on the level of the wanted signal is available in the digital domain, after decimation. Alterna-
tively, it can be derived from the modulation depth of the bitstream signal at the output of the ADC [122].
Table 9.4: Simulated SNR and contribution of quantization and of thermal

noise
Rin vin,max SQNR SNRth overall SNR
100k 0.35Vr ms 80dB 71.5dB 71dB

10k 0.035Vr ms 80dB 65dB 65dB
1k 0.0035Vr ms 80dB 50dB 50dB
Note that only the wanted signal is used for the gain control. Event hough interferers
are present and even while they may be much stronger than the wanted signal, the inter-
ferers do not cause overload because the ADC has attenuation at these frequencies. This
assumes that the limit on the allowable interferer (see evaluation in the next section) level
is not exceeded. A similar assumption would be required for a conventional architecture
based on a cascade of filters, programmable gain sections and an ADC: the programmable
gain control can be based on the strength of only the wanted signal if the interferers have
been suppressed sufficiently in the preceding filter.
The programming of the input resistance does not interfere with the noise shaping
and therefore the SQNR of the ADC is the same in all three settings. The overall SNR
including circuit noise, however, does depend on the value of Rin . This is due to the
following reasons:
• the thermal noise of Rin changes per setting;
• the input-referred thermal noise contribution of the other blocks changes because
of a different transfer of these sources;
• the signal power to which the noise is referred to, changes per setting
(vIN = 0.35Vrms for Rin = 100k while vIN = 3.5mVrms for Rin = 1k).
Table 9.4 lists a summary of the expected SNR due to quantization noise only (SQNR),
due to thermal noise only (SNRth ) and the overall SNR including both contributions.
ADC circuits
In order to avoid overlap with previous discussions, this section does not aim at a sys-
tematic and complete analysis of all circuits. Instead, a few distinct items, specific to the
feedback implementation are highlighted.
Noise budget: Despite of the efforts to increase the unity-gain frequency of the first
integrator, its gain at the edge of the signal bandwidth is still moderate (i.e. the unity-gain
Figure 9.20: Input-referred noise density of the implemented feedback ADC with
dominant contributions
frequency of the first integrator equals 1.9MHz while the signal bandwidth is 1MHz). As
a consequence, the second and the third stage of the loop filter contribute to the overall
noise of the ADC (see fig. 9.20). This was not the case in the feed forward design. Hence,
here, all integrators (except for the first one consuming 500µ A) are biased at 200µA
while in the feed forward ADC they were biased at 100µA. For use in section 6.3.3, it is
derived that the implementation factor of the feedback design, as compared to an equiva-
lent feed forward design equals: F = (500 + 4 · 200)/(500 + 4 · 100) ∼ = 1.5
Distortion at internal nodes: In the feed forward reference ADC, the transconductors
of the input stage represent the dominant source of distortion. In a feedback implemen-
tation, the consequent stages (i.e. the second stage, the third stage, etc.) also contribute.
This is due to the following reasons:
• the effect of distortion at these nodes on the overall distortion is larger than in the
feed forward implementation because the preceding loop gain is lower;
• the amount of distortion at the internal nodes is larger than in the feed forward case
because of a larger signal swing.
The latter has been illustrated in figures 6.15 and 6.16 in section 6.3 for a 3r d -order con-
ventional feedback ADC. However, the local feed forward path in the modified feedback
implementation changes the transfer function from the input of the ADC to the output of
Figure 9.21: Transfer function from the input of the ADC to the internal nodes
based on a linearized model of the modified feedback ADC
the first integrator. This is shown in fig. 9.21. In view of distortion, the signal swing at
the output of the first stage is most important. The frequency dependence of this signal
swing results in a strong frequency dependence of the IM3 -distortion of this ADC as well.
For example, in the wanted channel (i.e. from DC to 1MHz) the following frequency
dependence is expected:
• at the output of the first stage, the IM3 -distortion as a function of frequency has
slope +2 because the fundamental signal has slope +1 (see eq. B.2);
• the input-referred IM3 shows a slope of +3 because the transfer function from the
output of the first stage to the input of the ADC has slope +1 (i.e. the inverse of the
first integrator).
The frequency dependence of the input-referred IM3 for adjacent or far-off interferers can
be anticipated in the same way.
Graceful degradation: In case a large input signal is applied to a ADC, the

internal signals grow and the integrators may saturate. In case of a feed forward imple-
mentation, the forward paths by-pass the saturated stages and keep the loop functional.
This is referred to as “graceful degradation”. Even though the SQNR drops dramatically,
the loop remains operational and is able to recover fast when the input signal becomes
smaller again.
In a feedback implementation, this does not happen: the saturated integrators block
the loop operation because they are not by-passed. For the same reason, it takes a long
Figure 9.22: Implementation of “graceful degradation” (a) and transfer func-

tion with output voltage as a variable (b)
time to recover from an over-load condition. Therefore, an alternative implementation

of the “graceful degradation” technique is required. Fig. 9.22.a shows the implemented
solution: two diodes (M1 and M2) with reverse polarization are connected between the
outputs of the integrator. The threshold voltage of the diode transistors is about 250mV.
If the output swing exceeds this value, the diodes become a low-impedance load for the
gm C-integrator. Instead of saturating, the gm C-stage starts acting as a low-gain amplifier
(see fig. 9.22.b). Therefore, the order of the loop filter and of the noise-shaping reduces
while the modulator remains stable. As soon as the signal swing becomes smaller again,
the diodes go in the “off”-condition and the integration is resumed.
9.3.2 Evaluation
This ADC has been embedded in a test IC similar to the previous one; i.e. it has on-chip
clock generation, a bandgap reference, etc. The set-up for the evaluation is the same as
discussed before. First, the conventional performance parameters are evaluated. Next, the
interferer immunity is discussed.
Figure 9.23: SNR versus input swing for the three settings of Rin
Fig. 9.23 shows the SNR as a function of the digital output for the three settings of the
input resistance. Because of the programmable gain, the input-referred DR has extended
to 89.5dB. The peak-SNR remains moderate because of the simplified loop filter topology.
This results in relaxed specifications on the consequent blocks in the conditioning channel
and on a number of reference circuits (see section 9.5). Note that the peak-SNR and the
DR range differ per setting. This is due to the fact that the relative contribution of thermal
noise -compared to quantization noise- differs for the various Rin settings (see table 9.4).
This is also shown in the output spectra of fig. 9.24.a and b: for Rin = 100k the
shaping of the quantization noise is clearly visible. For Rin = 1k it is covered by an
almost white noise floor due to the thermal noise. This white noise floor even hides the
“bump” of quantization noise that is present in the image channel in fig. 9.24.a.
Fig. 9.25 shows the IM3 distortion in the wanted channels as a function of frequency.
Two nearby signals, i.e. at frequency f and f+30kHz are applied and the frequency f is
swept. As explained above, in the feedback ADC, dominant distortion occurs es-
pecially at the output of the first stage and, within the wanted channel, has a slope of
18dB/oct as a function of frequency (see discussion on page 161). Finally, note that the
measured IM3 is very similar for all settings. Even though the input voltage differs in
the three settings, the input current is always the same because Rin is scaled. By conse-
quence, also the internal signal swings -determining the distortion- are the same.
Table 9.5 gives an overview of the conventional ADC performance parameters. Eval-
uating the FOMSINAD of eq. 4.4 seems to indicate that the ADC does not perform well,
Figure 9.24: Full -scale input for Rin = 100 and for Rin = 1k
Figure 9.25: IM3 of wanted channels over frequency
in any of the settings, as compared to the ADCs listed in the appendix A. This demon-
strates a shortcoming of the use of FOMs: they are only suited for the assessment of a
stand-alone ADC. Here, the ADC includes signal-conditioning; i.e. it handles a very large
input-referred DR and, at the output, only provides a moderate SINAD. Therefore, the
ADC is also evaluated according to the FOMDR of eq. 4.1 where the input-referred DR is
filled out. As such, the performance can be compared favorably to e.g. the fundamental
limit of 1.6 × 10−20 J.
Interferer immunity
The evaluation of the interferer immunity concentrates on the setting where Rin = 1k
and the maximum wanted signal equals 3.5mVr ms . The reason is that, especially in the
case where the wanted signal is small, the interferers are likely to be larger. For the other
settings, the results for nearby interferers are similar or even better. The allowable input
level of far-off interferers, for these settings, is limited to the supply voltage because of
reliability issues.
Table 9.5: Measured performance of the feedback ADC (conventional para-

meters)
bandwidth 0-1MHz
sample rate 64MHz
Rin 100k 10k 1k
full-scale, diff. input 0.35Vr ms 0.035Vr ms 0.0035Vr ms
DR 71dB 64.5dB 49.5dB
peak-SNR 68.5dB 63dB 49dB
IM3 at 500kHz < −66dBc < −63dBc < −61dBc
FOMSINAD eq. 4.4 7 × 10−16 J 2 × 10−15 J 6 × 10−14 J
overall FOMDR eq. 4.1 5 × 10−18 J

Iquiescent 2.6mA @1.8V
active area 0.2mm 2
Aliasing limit on interferer immunity: Fig. 9.26 shows the result of an aliasing test
(for Rin = 1k) with an input of 64.37MHz at 0.02V and at 0.2V. The latter level is over
30dB higher than the full-scale level for wanted signals. The aliasing suppression equals
80dB here. For Rin = 10k it equals 90dB. These values are slightly lower than what is
simulated (probably parasitic mixing is more important than aliasing). For Rin = 100k
the alias suppression is over 100dB: the maximum input level is limited by the supply,
the corresponding alias is hidden in the noise. As predicted in section 6.3 the aliasing
suppression of this feedback ADC is significantly better than that of the feed forward
reference ADC (even while the latter has a 5th -order loop filter). This is discussed in
more detail in section 9.5.
Stable input range for interferers: In fig. 9.27 the measured stable input range is
compared to the simulated graph3 . Again, the simulation assumes small-signal operation
and, by consequence, only gives a coarse prediction near full-scale. Transient simulations
are much more accurate but are very time-consuming. Still, the slope of the curve is
predicted well and also the absolute accuracy of the prediction is acceptable.
3 Note that the simulated curve slightly differs from that in fig. 9.18. The latter was based on simulations of
an ideal model. Here, simulations on the implemented loop filter are used and, moreover, the various coefficients
have been slightly tuned to prevent overshoot in the adjacent channels.
Figure 9.26: Aliasing when applying an interferer at 64.37MHz at various input

levels (Rin = 1k)

Figure 9.28: Spurious responses when applying an interferer at 31.3MHz at var-

ious input levels (Rin = 1k)
Note that the curve is only depicted for positive frequencies. Contrary to fig. 9.14 for
the feed forward reference ADC, here, the stable input range is more or less symmetrical
around DC. This may be due to the order of the loop filter: the phase of the loop gain
changes more gradually in the present 4th -order design.
Spurious responses due to interferers: An interferer at 31.3MHz is applied and

its amplitude is increased until spurious components arise in the wanted channel (see
fig. 9.28). For Rin = 1k an interferer of 0.125V may be applied at this frequency without
affecting the DR listed in table 9.5. For the other settings, interferers at this frequency are
only limited by the supply voltage. It can be concluded that the feedback ADC suffers
much less from spurious responses than the feed forward ADC.
IM3 -distortion of interferers: The IM3 -distortion of the presented ADC strongly
varies with the frequency of the input signals. Especially when adjacent interferers are
applied, the internal signal swings may become large (see fig. 9.21) and cause distortion.
Applying two adjacent interferers at frequencies f 1 and f 2 such that 2 f 1 − f 2 = 500k H z
yields the graph of fig. 9.29.a showing IM3 as a function of f 1 . The input level is adapted
such that the interferers add up to the corresponding stable input range.
For Rin = 100k and Rin = 10k, note the following:
• the IM3 -distortion is similar for both settings: although the input voltage is different
in both cases, the input current is the same. By consequence, the internal signal
swings and the resulting distortion is similar as well;
Figure 9.29: IM3 distortion as a function of the input frequency for adjacent
interferers (a) and example for far-off interferers (Rin = 10k) (b)
• the IM3 -distortion changes with frequency. This is due to the frequency-dependent
signal swing at the output of the 1st integrator. From fig. 9.21, this is the node with
the largest signal swing (for adjacent interferers) and with the dominant distortion
component.
The curves of fig. 9.29.a have been verified in simulation by applying an equivalent dis-
tortion source at the output of the first integrator. This model is also used to calculate the
consequent limit on the allowable interferer level (see fig. 9.30). For Rin = 1k the graph
is different. This may be due to the following reasons:
• in this setting, the noise is significantly higher and partly adds to the distortion
energy resulting in an inaccurate measurement. This is also the reason why IM3 is
not shown beyond 2MHz4 ;
• the non-linearity of the input transconductors of the first stage becomes more impor-
tant in this setting because of the low value for Rin (see eq. 4.12). By consequence,
IM3 is higher and less dependent on frequency than in the other modes.
It should be noted that the measured IM3 -distortion of adjacent interferers is significantly
worse than that in the feed forward ADC. The main reason is the fact that, here, the entire
signal is fed into the loop via nested feedback while in the feed forward implementation
only the error signal is applied to the loop filter.
On the contrary, intermodulation is not a problem for far-off interferers (the internal
signal swings are small at these frequencies). This is illustrated with the two-tone test
of fig. 9.29.b: inputs at 28MHz and at 55.5MHz are applied at an amplitude of 0.5V for
Rin = 10k 5 . Note that this level is 20dB higher than the full-scale input level for wanted
4 At these frequencies, the stable input range is smaller than in the wanted channel. Hence, for the IM -test
3
the interferers must be applied with a smaller amplitude as well and the resulting IM3 -component gets hidden
in the noise.
5 This test is shown for R = 10k, because for R = 1k, intermodulation components are hidden in the
in in
noise.
(Rin = 1k)
signals for this gain setting. Even then, the resulting IM3 component at 500kHz is at
−73dBFS . This is sufficiently low in order not to affect the DR listed in table 9.5.
The above evaluation is summarized in the graph of fig. 9.30 for Rin = 1k. The
limitation due to IM3 -distortion is depicted for a target IM3 -value of −50dBc . This value
is chosen in order to demonstrate that, contrary to the other conditioning ADCs evaluated
in this chapter, in the present design, IM3 -distortion may yield the dominant limitation on
the allowable interferer level at adjacent frequencies. This depends on the target perfor-
mance. In order to improve on this limit, further down-scaling of the unity-gain frequency
of the first integrator is needed or more linear active stages must be designed. Both
approaches will increase the power consumption of the ADC though. At higher frequen-
cies, the allowable interferer level is set by the stable input range of the ADC and for
even higher frequencies, the supply voltage is limiting the levels. Interferers applied near
32MHz must remain smaller than +25dBFS in view of spurious responses. Interferers ap-
plied near 64MHz are limited to +28dBFS in view of aliasing. For Rin = 10k and even
more for Rin = 100k, especially the supply limit becomes more restricting.
9.4 FFB-ADC
A FFB-ADC is presented as a third implementation of the conditioning channel for the
Bluetooth receiver. This design is an illustration of the theory developed in section 6.4.
The ADC combines the advantages of the feed forward and the feedback topologies pre-
sented in the previous sections; i.e. it features both excellent power/performance and a
high immunity to interferers.
9.4. FFB-ADC 171
9.4.1 Design
In an analogy to the previous section, the discussion on the ADC design covers some
architectural choices (a.o. the choice of the loop filter and of the added filters), motivates
the integration of programmable gain and analyzes some implementation related aspects.
ADC architecture
The FFB-ADC is constructed from a conventional ADC with a 4th -order, real, feed for-
ward loop filter H (s). (This ADC will be referred to as the “original ADC”.) In fact,
the 5th -order, complex loop filter of the ADC in section 9.2, could have been used, just
as well. However, a loop filter with real coefficients is preferred because it results in a
lower complexity of the design, layout and evaluation. Moreover, the simulation time for
a transient analysis reduces by more than a factor of two compared to a complex design.
These arguments motivate the use of a “lower risk” loop filter with real coefficients in
order to concentrate on the true innovation of this design; i.e. on the unrestricted filtering
in the STF. Furthermore, the choice for a design with real coefficients has the following
consequences:
• for use in a quadrature receiver, two ADCs need to operate in parallel in order to
digitize both the I and the Q signals;
• the effective over-sampling in a ADC with real coefficients is only half of that
when using a complex loop filter. At this low over-sample factor, a 5th -order loop
filter doesn’t allow better noise shaping (because of stability) than a 4th -order filter.
While in the complex designs, cross-coupling paths between quadrature signals
were used to implement a resonance and to provide additional gain in the wanted
channel. This is not possible in a design with real coefficients. For these two
reasons, the SQNR of this ADC is only 74dB.
It is shown later on that, because of the filtering behavior of this ADC, programmable
gain control can be added and, then, a much larger input-referred DR becomes available.
While the choice of the loop filter determines the SQNR and the NTF of the ADC,
the STF can be designed independently, e.g. by adding complementary filters HLPF and
HHPF (see section 6.4). In the presented design, the parallel configuration of fig. 6.19.a is
used as this results in a favorable analog implementation. Likewise, for HLPF and HHPF
a simple 1st -order implementation is used. Fig 9.31 shows the STF of the 4th-order feed
forward ADC without the complementary filters and the effect of adding the filters HLPF
and HHPF . These filters are implemented with a −3dB-frequency of 3MHz. This choice
results as a compromise between:
• good attenuation of adjacent interferers and of the overshoot that is present in the
STF of the original feed forward ADC;
• limited droop and group delay variation in the wanted channel.
The latter requirements are application dependent and are outside of the present scope.
In general, the STF of the FFB-ADC is optimized by choosing the appropriate charac-
Figure 9.31: STF of the FFB-ADC with unrestricted filtering
teristic for HLPF and HHPF . On the contrary, the loop filter H (s) should not be modified
compared to that of the original ADC as it has been designed for optimal noise shaping.
Integration of programmable gain

As argued for the feedback ADC on page 159, a filtering STF is optimally exploited in
combination with programmable gain control of the input signal. Since the ADC is highly
immune to interferers, its input range can be scaled according to the strength of the incom-
ing wanted signal. As such, the input-referred DR is extended significantly. Moreover,
the merge of programmable gain and filtering into a ADC yields a true equivalent of
a conventional conditioning channel.
The gain programming is implemented in the same way as was done for the feed-
back ADC: the input resistance of the ADC is switched between 1k, 10k and 100k
in order to, respectively, accommodate wanted signals up to 5mV, 50mV and 500mV
of amplitude. In addition, in the present design, the quiescent current of the first stage
is adapted along with the input resistance6 . This is motivated further on as one of the
implementation aspects.
Implementation aspects
Fig. 9.32 shows a block diagram of the implementation. Compared to fig. 6.19.a, the
first stage H1 (s) of the loop filter H (s) is shifted in front of the summation point and is
6 The scaling of the quiescent current can be implemented in the feedback design as well and, likewise, would
improve the average power/performance ratio.
9.4. FFB-ADC 173
Figure 9.32: Block diagram of the implemented FFB-ADC
duplicated in the parallel feedback path. Shifting H1 (s) before HLPF and HHPF yields the
advantage of an easier implementation for these filters:
• HLPF is in between two active stages (acting as buffers) and a passive implementa-
tion becomes possible;
• the series connection of the integrator H1 (s) and the 1st -order high-pass filter HHPF
can be replaced by a low-pass filter, resulting in an easier implementation.
Fig. 9.33 shows the block diagram of the actual implementation. The 4th -order loop fil-
ter is implemented in a similar way as in the previous designs. The resonance around
the 4th integrator provides some additional gain at the edge of the wanted channel. The
combination of the resistor R f and the subsequent capacitor C1 constitutes HLPF . (For
simplicity, a floating capacitor of value C1 /2 is drawn. The actual implementation uses
two capacitors of value C1 that are referenced to ground.)
In the parallel feedback path, the equivalent low-pass filter is built around OTA g f b .
The combination of R f b and C1 realizes H1 (s) and the combination of R f and C1 real-
izes HHPF . Note that, here, an active implementation is preferred over a passive one: the
active stage supplies the signal-dependent current to the integration capacitor such that
the DAC reference has a constant load. A passive implementation would pollute the DAC
reference that is being shared with other circuits, with a signal-dependent load.
Finally, in order to perform the summation of the parallel path in the current domain,
two additional transconductors fb1 and fb2 are needed. Their quiescent current of 10µA
is negligible.
Internal signal swings: When discussing the feedback topology of section 9.3, the
signal swing at the output of the first integrator was identified as an important problem.
It is the dominant source of distortion and limits the unity-gain frequency -and thus the
gain- of the first integrator. Here, the addition of the filters also affects the signal swing at
this node.
Figure 9.33: 4th -order ADC with unrestricted filtering
Compared to the feed forward ADC of section 9.2, the signal swing has increased in
the wanted channel (see fig. 9.34). However, the maximum signal swing remains compa-
rable. In addition, the unity-gain frequency of the first integrator (fug = 2.9MHz) is about
equal to that of feed forward ADC of section 9.2 (fug = 2.8MHz).
Compared to the feedback implementation of section 9.3, though, the signal remains
significantly smaller, even while the unity-gain frequency of the first integrator is 1.5
times larger here.
Noise and distortion of the added filters: The major noise contributions of the design
are shown in fig. 9.35 that is discussed further on. Since H1 has been shifted in front, the
noise and distortion of HLPF is counteracted by the preceding gain. The same is true is for
OTA g f b in the feedback path: when referring to the input, its contribution to noise and
distortion is suppressed by H1 . This OTA is a scaled copy of gm1 : while gm1 consumes
500µA, g f b can be biased at only 50µA because of the above reasons. For use in sec-
tion 6.4.3 it is mentioned that the implementation factor of the FFB-ADC, as compared
to the original ADC, equals F ∼ = 1.1.
It can be concluded that the added filters hardly increase the overall noise, distortion
or current consumption of the original ADC. The basic reason is the fact that the added
filters are put inside of a closed-loop system. This is a key difference compared to the
conventional cascade of an analog filter and an ADC. There, the filter would be in the
signal path and, consequently, its noise and distortion would contribute.
Programmable quiescent current for the first stage: As in the previous ADCs, the
first stage of the loop filter contributes dominantly to the overall circuit noise. There-
9.4. FFB-ADC 175
Figure 9.34: Transfer function from the input of the ADC to the output of the first
integrator based on a linearized model
fore, its nominal quiescent current is set at 500µA (just as in the input stage of previous
ADCs). A small improvement of the average consumption is implemented by adapting
this current along with the programming of Rin to the magnitude of the wanted signal. As
such, the quiescent current of the first stage is reduced to 200µA when Rin is switched to
100k because a large input signal is present.
Fig 9.35 shows the simulated, input-referred noise-density and the major contribu-
tions. For Rin = 1k (and for Rin = 10k) the first integrator is dominant. For Rin = 100k
the contribution of gm2 and of g f b become important because the preceding gain drops
(the gain of the input integrator is inversely proportional to Rin ). In general terms, the
effect of the programming of Rin on the various noise contributions is similar to what has
been discussed for the feedback ADC on page 159.
Mismatch between HLPF and HHPF : A mismatch between the time-constants of the
added filters HLPF and HHPF would violate the complementarity posted in eq. 6.19. The-
oretically, this affects the stability of the loop and an additional phase margin must then
be taken into account during the design phase. For common mismatch values, though,
this issue is not at all restricting. For example, it can be calculated that a 5%-mismatch
on the time-constants causes a phase shift of less than 0.3 degrees at half the sample rate.
Spread on HLPF and HHPF : Spread on the target −3dB-frequency of the filters results
in spread on the STF of the ADC and thus on the filtering of the interferers. If this spread
cannot be tolerated, this problem must be dealt with as in any continuous-time filter; i.e.
either the time-constants are calibrated or a margin on the filtering is implemented.
Figure 9.35: Input-referred circuit noise for Rin = 1k (a) and Rin = 100k (b)
9.4.2 Evaluation
Similar to the ADCs of section 9.2 and 9.3, this design is embedded in a test-IC including
a bandgap reference and on-chip clock generation. The evaluation and the discussion
hereon are conducted in a similar way as well.
Fig 9.36 shows the measured SNR as a function of the input voltage for the various values
of Rin . Adapting the value of Rin depending on the magnitude of the wanted channel re-
sults in an input-referred DR of 89dB. In addition, the output-referred peak-SNR remains
moderate. It is discussed further on (section 9.5) that this is beneficial on a system level.
As for the feedback ADC, the difference in peak-SNR between the various settings is due
to a different relative contribution of circuit noise and quantization noise.
Fig. 9.37 shows the output spectrum of an IM3 -test in the wanted channel. For
Rin = 100k, IM2 - and HD2 -components are visible. These have been traced back to the
generator7 . In addition, some shaping of quantization noise is visible within the wanted
channel. For Rin = 1k, circuit noise is by far dominant and the noise-density corre-
sponds to the curve simulated in fig. 9.35.a. In general, IM3 < −60dBc for Rin = 100k
and Rin = 10k. For Rin = 1k, it is below −50dBc . Qualitatively, this increase is pre-
dicted from eq. 4.12 and considering the values for vin , Rin and gm . Quantitatively, the
difference deviates from what is predicted by eq. 4.12. Compared to the IM3 -distortion for
Rin = 100k, eq. 4.12 predicts 4dB lower and 16dB higher distortion for Rin = 10k and
Rin = 1k respectively. A possible explanation may be in the fact that the input transistors
operate in weak (instead of strong) inversion or in the fact that other distortion sources
contribute as well.
7 For most other test, generator induced spurious are suppressed by calibration before measuring.
9.4. FFB-ADC 177
Figure 9.36: SNR versus input swing for the three settings of Rin
Figure 9.37: Two-tone test in the wanted channel evaluating IM3 -distortion for
Rin = 100k(a) and for Rin = 1k (b)
In the wanted channel, the measured IM3 -distortion is more or less independent of
frequency. Hence, most likely, it is due to a non-linearity in the input stage or in the feed-
back path.
Table 9.6 gives on overview of the measured conventional performance parameters.

In addition, it lists the achieved FOMSINAD according to eq. 4.4. Again, this FOM is not
suitable for the evaluation of conditioning ADCs because these limit the SINAD towards
the output. Therefore, eq. 4.1 is evaluated as well. It benchmarks the ADC according to
the input-referred DR it provides. This table lists the performance of one ADC with real
Table 9.6: Measured performance of one real, FFB-ADC (conventional para-

meters)
bandwidth 0-1MHz
sample rate 64MHz
Nyquist sample rate 2MHz
Rin 100k 10k 1k
full-scale, diff. input 0.35Vr ms 0.035Vr ms 0.0035Vr ms
DR 65dB 59dB 49dB
peak-SNR 59dB 57dB 46dB
IM3 (in wanted channel) < −60dBc < −60dBc < −50dBc
Iquiescent 0.85mA 0.85mA 1.15mA
FOMSINAD eq. 4.4 2 × 10−15 J 3 × 10−15 J 5 × 10−14 J
overall FOMDR eq. 4.1 2.6 × 10−18 J

coefficients. For comparison with the complex implementations of the previous sections,
the performance, area and power consumption of two parallel ADCs should be consid-
ered. Then, in the quadrature configuration, a 3dB higher DR and peak-SNR are achieved
and area and power consumption double.
Interferer immunity
Fig. 9.38 illustrates the operation of the presented FFB-ADC. A very small wanted input
is applied together with two much larger, far-off interferers. The latter are applied at an
input level corresponding to half the stable input for that frequency. At the output, the
wanted signal appears near the full-scale output level. Relatively, the interferers, have
been attenuated. This demonstrates the filtering STF of the presented ADC. In addition,
this measurement also proves the linearity with respect to far-off interferers. Even though
the interferers are applied at an input level that is significantly higher than the wanted
channel ( f 2 is applied 15dB higher, f 3 is applied 24dB higher), at the output, their
IM3 -distortion is 55dB below the wanted channel.
The measurement of fig. 9.38 gives a first impression of the interferer immunity of
this ADC. A systematic evaluation of the various limitations on the allowable interferer
level is presented below. Afterwards, the most restricting limitations are summarized.
9.4. FFB-ADC 179
Figure 9.38: Tri-tone input (a) and output spectrum (b) demonstrating filtering
and linearity ( f 1 = 700k H z, f 2 = 4.8M H z and f 3 = 10M H z;
Rin = 1k)
Aliasing limit on interferer immunity: Alias suppression of 69dB, 67dB, and 76dB
is measured for Rin equaling 100k, 10k and 1k respectively. These values are lower
than what is expected from eq. 6.3. Probably, parasitic mixing with the clock frequency
is dominant over aliasing. This hypothesis, would also explain the difference in the
measured values.
The alias suppression is better than that of the feed forward ADC of section 9.2
(60dB of suppression is achieved there) but worse than that of the feedback ADC of
section 9.3 (achieving over 80dB of alias suppression). This is explained by the difference
in the order of L 0 and of L 1 . This is discussed in more detail in section 9.5.
Stable input range for interferers: Fig. 9.39 compares the measured stable input to
the simulated level based on a linearized small-signal model for Rin = 1k. For adjacent
frequencies, the prediction from the model may deviate from the measurement by a few
decibels. For far-off frequencies, the prediction of the absolute value and of the slope
of the curve is very good. Measurements over several samples and also for Rin = 10k,
show that the maximum deviation between the simulated and the measured overload
level is about 2dB.
Spurious responses due to interferers: In this design, the order of L 0 equals 2 near
half the sample frequency. Compared to the feed forward ADC of section 9.2 (with L 0 of
order 1 in this frequency range), the interferer is attenuated stronger before being applied
to the quantizer. Since the quantizer represents a “strong non-linearity”, this results in
a major improvement in terms of spurious responses. Following a similar reasoning,
this ADC must be somewhat less robust, in this respect, than the feedback ADC of
section 9.3 having L 0 of order 3 near half the sample frequency. This is demonstrated
by the measurement in fig. 9.40. For Rin = 1k, interferers up to 100mV can be applied
without affecting the DR in the wanted channel. Note, the full-scale input for wanted
Figure 9.40: Spurious responses when applying an interferer at 31.3MHz at var-

ious input levels (Rin = 1k)
channels is only 5mV in this setting. For Rin = 10k and Rin = 100k, the allowable
interferer level near 32MHz is limited by the supply voltage.
IM3 -distortion of interferers: For Rin = 1k, fig. 9.41 shows IM3 -distortion over
frequency. In this measurement, interferers at frequency f 1 and frequency f 2 are
applied at an input level corresponding to half the stable input range for that frequency
(see fig. 9.39). These frequencies are chosen such that 2f1 −f2 falls in the wanted
9.5. EVALUATION OF THE CHANNELS 181
Figure 9.41: Measured IM3 -distortion of interferers for Rin = 1k
channel. Within the measured frequency range, IM3 -distortion decreases as a function
of frequency while the input level of the interferers increases according to the curve of
fig. 9.39. Hence, the distorting node may be inside of the loop. This is not investigated
further because, nor for adjacent interferers, nor for far-off interferers, IM3 -distortion is a
restricting limit on the allowable interferer level. This becomes clear from the discussion
on fig. 9.42 below. The same conclusion holds for Rin = 10k and Rin = 100k.
Fig. 9.42 gives a summary of the limitations on the allowable interferer level over
frequency for Rin = 1k. For most frequencies, the stable input range causes the
restricting limit. Near 32MHz and near 64MHz the interferers must be smaller than
+23dBFS and +24dBFS respectively to keep spurious responses and alias components
low enough. IM3 -distortion is not a restricting limit for this ADC For Rin = 10k and
Rin = 100k, the graph is similar except for the far-off interferers being limited by the
supply voltage.
9.5 Evaluation of the channels

In this chapter, three solutions for the signal conditioning channel in a highly-digitized
Bluetooth receiver have been presented. First, in section 9.5.1, the presented ADCs are
briefly benchmarked to published designs with a similar bandwidth. Next, in section 9.5.2,
the ADCs are compared against each other. Since all have been implemented in the same
technology and -to a large extent- reuse the same circuits, this is a fair comparison. Fi-
nally, the presented concept of digitizing the channel by integrating filtering and program-
mable gain into the ADC, is evaluated by comparison with highly-analog channels and
with channels using “conventional digitization”.
9.5.1 Benchmark with published ADCs

Only for the reference ADC of section 9.2, this benchmark is straightforward. It can be
evaluated using the FOMSINAD of eq. 4.4 and compared to other ADCs, a.o. [33], [123],
[124], [39], [28] and [125], published for this application area and listed in table A.1.
Together with [28], it achieves the best FOM among the converters targeting a bandwidth
of 1MHz.
The conditioning ADCs of sections 9.3 and 9.4 cannot be benchmarked in this man-
ner because they target to reduce the SINAD towards the output. Instead, the FOMDR of
eq. 4.1 can be used to evaluate their input-referred DR. In this aspect they outperform
“conventional ADCs” by two orders of magnitude, typically; i.e. FOMDR is of the order
of 10−19 J for the conditioning ADCs while it is ∼ 10−16 J for state-of-the-art “conven-
tional ADCs”. This demonstrates that a major power saving is possible when targeting
optimization of the overall channel instead of optimizing only building blocks.
9.5.2 Comparison of the presented ADCs

The ADCs of sections 9.2, 9.3 and 9.4 are compared to each other. For simplicity, they
are referred to, here, as the feed forward ADC, the feedback ADC and the FFB-ADC
respectively. Table 9.7 gives an overview of the conventional performance parameters
for the three ADCs. The figures in the last column refer to the quadrature configuration
of two FFB-ADCs. Since three settings are possible for the operation of the feedback
and the FFB-ADCs, their performance can be summarized using the combination of the
input-referred DR and the SNR at the output. The discussion on this table distinguishes
between conventional performance parameters and interferer immunity.
Table 9.7: Comparison of the ADCs in terms of conventional parameters
feed forward feedback 2 filtering

ADC ADC ADCs
(sec. 9.2) (sec. 9.3) (sec. 9.4)
active area 0.22mm 2 0.2mm 2 0.27mm 2

power consumption 4.4mW 4.7mW <4.1mW
input-referred DR 76dB 89.5dB 92dB
peak-SNR at output 75.5dB 68.5dB 62dB
FOMDR eq. 4.1 1 × 10−16 J 5.3 × 10−18 J 2.6 × 10−18 J
accuracy of references 0.01% 0.2% 0.2%
Comparison of conventional performance parameters: All solutions have a compa-

rable power consumption and occupy a similar chip area. Two observations stand out.
First, the solutions differ in the fact that both the feedback ADC and the FFB-ADC
have a significantly larger input-referred DR and a much moderate peak-SNR at the out-
put than the feed forward reference ADC. This advantage is obtained from the architec-
tural innovation of integrating explicit signal conditioning into the ADC. The conditioning
channel as a whole becomes more efficient in terms of power/performance:
• FOMDR of the ADC improves by two orders of magnitude: for a fixed power con-
sumption and maximum signal level, a larger input-referred DR is achieved. IM3
distortion is lower though. In fact, all performance figures evolve in a direction that
matches the channel requirements: a large DR is highly desirable, on the contrary,
IM3 -distortion of the wanted channel can be as high as −18d Bc without affecting
the bit error rate.
• A power and area reduction for the analog reference circuits: since the DAC in the
feedback path of the ADC can have a lower peak-SNR, the reference circuits for the
DAC can have about 20 times lower accuracy as well. The required accuracy8 is
listed in the last row of table 9.7. This is the accuracy needed to maintain the full
DR for Rin = 1k and listed in tables 9.5 and 9.6.
8 These numbers only give a first-order indication. The allowed “noise” depends on the frequency spectrum,
the stochastic distribution and on the shape of the DAC output pulses [126].
• Some power and area reduction in the following blocks of the channel: the decima-
tion filter can be dimensioned to a smaller peak-SNR, reducing the value of ENOB
in eq. 4.19. The improvement in peak-SNR, as listed in table 9.7, is 7dB for the
feedback ADC and 13.5dB for the FFB-ADC. From eq. 4.19, the power consump-
tion of the decimation filter may then reduce by 20% up to 40%9 . The bit rate in
between digital blocks (e.g. over an inter-die interface) is reduced as well. This
saving can be derived from eq. 7.4 but is minor here.
In general, the smaller peak-SNR at the output may also allow a lower sample rate for the
ADC (here, only the order of the loop filter has been changed). This would further
simplify the clock generation (next to the milder accuracy requirement) and reduce the
input sample rate of the decimation filter (next to the lower ENOB requirement).
Second, the FFB-ADC achieves a 2-times better FOM than the feedback ADC. This
reflects the discussion in section 9.3 on the difficult balancing of power/performance. In
the feedback ADC, noise and distortion are balanced by tuning coefficients and tweaking
quiescent currents of internal stages. On the contrary, the FFB-ADC is designed from
a topology that -inherently- achieves a better power/performance because of the overall
feedback.
Comparison of interferer immunity: This comparison is conducted based on the

graphs in fig. 9.43. For clarification of the graphs, first note the following.
On the y-axis, the antenna-referred signal power is plotted (instead of the input sig-
nal to the ADC) in order to easily link with specifications from the Bluetooth standard
[112]. For example, the maximum specified input power equals −20dBm. The gain of the
receiver front-end must then be designed such that this signal is applied at the full-scale
input of the ADC.
The dotted line represents the envelope curve of the specified maximum input sig-
nals over frequency. Note, the standard does not require that the receiver simultaneously
accommodates these maximum inputs.
The solid line indicates the allowable input level over frequency as introduced in
fig. 9.16 (positive frequencies only) and fig. 9.30. At the bottom end of the scale, the level
of the input referred noise of the ADC is indicated.
Hence, the interferer immunity of the feedback ADC of section 9.3 is -in first order-
comparable to that of the FFB-ADC of section 9.4. Therefore, it is left out of the first
part of the discussion, comparing the feed forward reference ADC to the FFB-ADC. Af-
terwards, the feedback ADC is briefly compared to the FFB-ADC.
The feed forward reference ADC of section 9.2 can accommodate all specified in-
terferer levels and provides sufficient DR. (A receiver sensitivity of −70dBm is required.
9 Depending on the required interferer immunity, more PGA-settings can be applied and the peak-SNR of the
ADC can be further reduced. This leads to a more significant saving in the decimation filter. The automatic gain
control becomes more difficult though.
Figure 9.43: Comparison of the dynamic range and interferer immunity (re-
ferred to the input power at the antenna) of the feed forward ref-
erence ADC (a) and of the quadrature configuration of two FFB-
ADCs (b)
At that input level, 18dB of SNR is needed for the digital demodulation. In addition, 8dB
margin is foreseen in view of the front-end noise). As such, the conventional feed forward
ADC enables Bluetooth certification of the receiver of fig. 9.2, even while no analog filters
or VGA are used. However, the margin on the interferer immunity is tight. In addition, the
de-facto sensitivity of published Bluetooth receivers is around −82dBm (a.o. [127], [115],
[116], [120], etc.) or lower (−88dBm is reported in [128]). In order to be competitive, the
present receiver needs more DR in the ADC or it needs preceding analog conditioning.
On the contrary, for the same peak power consumption, the FFB-ADC does enable
a Bluetooth receiver achieving the de-facto sensitivity goal and featuring a much more
robust operation in the field:
• For Rin = 100k, the basic Bluetooth specifications (with respect to noise and in-
terferer tests) are met. Still, little margin on the sensitivity is available.
• Switching Rin to a lower value improves the sensitivity (the antenna-referred noise
of the ADC is −112dBm for Rin = 1k).
• The inherent interferer immunity considerably improves the operation in the field
(with other or more combinations of interferers than those specified in the Bluetooth
standard) and the overall robustness of the receiver.
Unfortunately, the latter advantage -although extremely important- is difficult to quantify

uniquely. That would require a multitude of tests in order to evaluate all common combi-
nations of interferers at various amplitude levels of the wanted channel.
Figure 9.44: Allowable interferer level of the feedback ADC of section 9.3 com-
pared to that of the FFB-ADC of section 9.4
Alternatively, the concept of the conditioning ADC could be used to lower the
power consumption of the receiver while keeping the performance equal to that of the
conventional feed forward ADC.
Finally, the interferer immunity of the feedback ADC of section 9.3 is compared to
that of the FFB-ADC of section 9.4. This comparison is visualized in fig. 9.44. The
following is observed:
• For far-off interferers, the feedback ADC is more immune. This is especially due
to its larger stable input range at these frequencies and to a better alias-suppression.
Still, for many standards and in many receiver topologies, the immunity achieved
in the FFB-ADC may be well sufficient.
• For adjacent channels, both ADCs achieve a similar degree of immunity10 . The
feedback ADC attenuates adjacent channels more but its distortion limits the al-
lowable input to a level comparable to that of the FFB-ADC.
For the feedback ADC, the latter limit can only be improved at the expense of a higher
power consumption (see discussion on page 170). In the FFB-ADC, the allowable ad-
jacent interferer level is limited by the stable input range. It can be improved by imple-
menting a smaller f−3dB for the added filters or by changing the order of the filter. This
may have a noise penalty (the noise of a.o. the second stage of the loop filter becomes
more important) or a distortion penalty (the internal signal swings increase) respectively.
Still, fundamentally, more degrees-of-freedom are available for the optimization of the
FFB-ADC and therefore, the power penalty can be kept minimal.
10 This is true for the current example of an IM -target of −50d B and with the current implementation
3 c
parameters.
Table 9.8: Comparison of the ADCs for operation in a wireless system
feed forward feedback FFB-ADC

ADC (sec. 9.2) ADC (sec. 9.3) (sec. 9.4)
power/performance a + ++ +++
interferer immunity + ++ ++
opportunities for ? + ++
further optimization
a Assuming only a moderate SNR is needed for demodulation, a power reduction is obtained by reducing
the peak-SNR early on in the receive chain, i.e. at the output of the ADC. In other words, in these systems a
moderate peak-SNR (in combination with a large input DR) is favorable.
These degrees-of-freedom originate from the addition of the extra filters. This brings
us to another limitation of the feedback ADC; i.e. the NTF and STF cannot be optimized
independently. Designing an optimal NTF results in a fixed value for f−3dB of the STF.
In general, if more over-sampling is applied and optimal noise shaping is pursued, the
loop gain remains high over a wider frequency range. Consequently, f−3dB of the STF
increases.
The discussion is summarized in table 9.8. The feed forward ADC achieves a very
competitive power/performance ratio comparing to state-of-the-art. In addition, it is suffi-
ciently immune to interferers in order to enable a Bluetooth receiver without analog signal
conditioning. It remains an attractive solution for systems requiring a high SNR for the
digital processing. As an example, ADSL -using a more complex modulation scheme-
could be considered.
Most wireless communication systems, though, use a modulation scheme that needs
less SNR for demodulation. This is also the case for the present Bluetooth receiver. Here,
the feedback and the FFB-ADC enable a much better sensitivity and higher robustness.
Especially, the FFB-ADC is attractive. The present implementation performs slightly bet-
ter than the feedback ADC. More important, more degrees-of-freedom are available for
further improvements or for adaptation to other interferer spectra.
9.5.3 Benchmark with published Bluetooth conditioning channels

Publications on Bluetooth receivers hardly present data on the baseband part of the signal
conditioning channel separately. Hence, the RF part is included in the present benchmark.
The RF part of the low IF receiver (the receiver of fig. 9.2) consists of an LNA and a
passive mixer only. It achieves −71dBm of sensitivity while the peak power consumption
is 32mW [121]. Table 9.9 compares this receiver to an analog Bluetooth receiver [127]
Table 9.9: Benchmark of the Bluetooth receiver
an. con. + an. con. + dig. con. +

an. demod. dig. demod. dig. demod.
[127] [120] [121]+ ADC of sec. 9.2
sensitivity -82dBm -83dBm -71dBm

max. input 0dBm -5dBm -19dBma
P 75mW 58mW 32mW
area 4mm 2 ?mm 2 3.5mm 2
technology 0.18µm CMOS 0.13µm CMOS 0.18µm CMOS
a The maximum input power is smaller than expected. Probably, this is due to a problem in the matching
network at RF.
and to a receiver using analog signal conditioning but digital demodulation [120]. (The
latter receiver uses various sections of discrete-time filtering and a switched-capacitor
ADC.) To the author’s knowledge, these receivers are among the best reported. Note
that, as opposed to the various solutions presented in this chapter, all reported Bluetooth
receivers need analog signal conditioning.
A comparison of power/performance between the three receivers of table 9.9 is not
straightforward because of the difference in sensitivity. It requires extrapolation of the
power consumption of the receiver of the last column (i.e. of fig. 9.2) to a sensitivity of
−83dBm. Probably, the highest power increase is needed in the LNA and in the mixer
(likely the passive mixer must be replaced by an active solution to meet the performance
target). The PLL and the VCO can remain unaltered because their performance is believed
to be good enough. Most important, the ADC of section 9.2 can be replaced by the ADC
of section 9.3 or, preferably, by the ADC of section 9.4. This can be done without any
power penalty at all while the required DR is achieved with a margin of about 10dB. The
decimation filter that follows, becomes easier than in the present channel. On the other
hand, an automatic gain control algorithm must be implemented.
Although extrapolation of the power consumption in the RF-part is very speculative,
the feasibility of a much improved baseband, at constant area and power consumption,
has been proven in this chapter.
9.6 Conclusions
The theory on conditioning ADCs, as presented in chapter 6, is illustrated with design
examples for a Bluetooth receiver. Conclusions are listed below.
The feed forward reference ADC of section 9.2, with signal-conditioning in the deci-
mation filter, achieves a state-of-the-art FOMSINAD of 1×10−6 J and enables a Bluetooth-
compliant receiver without the need for analog signal conditioning.
FOMSINAD is not suited for benchmarking the ADCs of sections 9.3 and 9.4. Instead,
FOMDR is used to evaluate the input-referred DR: in this respect, the conditioning ADCs
surpass state-of-the-art by two orders of magnitude. This is the virtue of integrating sig-
nal conditioning into the ADC. Moreover, these ADCs allow a power saving in the
consecutive blocks of the channel by a factor of about 20.
The feed forward reference ADC with signal-conditioning in the decimation filter, is
recommended for receivers that benefit from a large SNR at the output. For receivers that
need only a moderate SNR (e.g. most wireless communication systems), a conditioning
ADC with a filtering STF is recommended.
The FFB-ADC -as compared to the feedback ADC- is capable of achieving a better
power/performance balance and features more degrees-of-freedom to optimize the inter-
ferer immunity.
Chapter 10
General conclusions
Signal conditioning channels are being digitized in a quest for flexibility and to benefit
from technology scaling. While in the signal processing arena digitization has lead to
a power saving per computation, this benefit is not self-evident for the conditioning
channel. In this context, this book aims at improving the power/performance relation of
the conditioning channel by balancing analog and digital signal conditioning.
Digitizing the conditioning channel puts a burden on the data converters. In receiver
applications, the bandwidth and linearity requirements on, especially, the digitized A/D
channel become very challenging due to the presence of interferers. It is shown that
a single-bit, continuous-time modulator with a feed forward loop filter promises
the best power/performance balance for the ADC. This topology is used as a reference
throughout the book.
Analyzing the power/performance relation of analog circuits, ADCs and digital
functions, indicates that a ADC may constitute a power-effective replacement for a
cascade of analog blocks. Moreover, it is clear that the power consumption of the ADC
becomes dominant in many conditioning channels, even when these are being digitized.
In view of the slow advances in the power/performance balance of ADCs -and analog
circuits in general- it is concluded that architectural innovation is needed to enable
power-effective digitization.
This conclusion is also supported by the analysis of a full-digital conditioning
channel: next to the performance challenge for the ADC and the analog reference
circuits, even the sample rate in the digital blocks becomes cumbersome. At present,
straightforward digitization -by shifting the ADC towards the antenna- is only feasible
for narrow-band systems. Even then, the linearity requirement remains very demanding.
As an alternative, the concept of “conditioning ADCs” is proposed: instead of
replacing analog conditioning circuits by digital functions, the signal-conditioning can
be integrated into the “broader ADC”. The key pillar, enabling the conditioning
ADC, is the fact that continuous-time ADCs are -to a large extent- immune to
interferers.
This characteristic is first exploited in a conditioning ADC with the signal-con-
ditioning in the decimation filter. The available decimation filtering and word-length
191
192 GENERAL CONCLUSIONS
scaling at the same time provides channel filtering and digital VGA. As compared to the
straightforward digitization, this solution reduces the bandwidth and sample rate related
problems.
When using a topology that features a filtering signal transfer function, it can
be combined with variable gain control of the input signal. The signal-conditioning
that is, as such, integrated into the loop, enables a large input-referred DR for the
ADC while, at the output, the SINAD remains moderate. Consequently, this solution
provides an additional advantage by relaxing the accuracy required of a.o. the DAC
and its references. In topologies with nested feedback, the achievable filtering of the
signal transfer function is restricted. In addition, the power/performance balance is less
attractive than that of the reference ADC. Therefore, a “filtering-feedback ADC”
is proposed as a better solution. It allows for unrestricted design of the signal transfer
function and, moreover, it is based on the same low-power topology as the reference ADC.
Appendix A
Overview of published ADCs
Table A.1 lists ADCs published between 2000 and 2004. It includes publications
at conferences such as ISSCC, CICC, ESSCIRC and the VLSI Circuits Symposium and
papers from the Journal of Solid-State Circuits. Notice:
• all ADCs are implemented in CMOS technology unless “BiC.”is mentioned to in-
dicate a BiCMOS implementation. In the second column the feature size of the
technology is listed;
• the FOM in the 6th and 7th column are that of eq. 4.4 and eq. 4.6, respectively.
The differences between and the use of these FOMs is discussed in section 4.2. It
is repeated here that the lower limit on FOMSINAD equals 1.6 × 10−20 J (assuming
only noise);
• m, L and N stand for the over-sample ratio, the order of the loop filter and the num-
ber of bits in the quantizer, respectively. In case, several numbers are mentioned
for L and N, this indicates that a cascaded topology-i.e. a MASH ADC- is
reported;
• “CT” and “DT” stand for continuous-time and discrete-time, respectively. “ff” and
“fb” indicate whether a loop filter with feed forward or feedback stability compen-
sation is used. If “feedin” is added, branches from the input of the ADC to internal
nodes of the loop filter are present, such as in the hybrid topology of section 6.3.4.
193
Table A.1: Survey of ADCs published from 2000 until 2004. 194
Ref. Techn BW SINAD P FOMSINAD FOMENOB m Type L N

µm [kHz] [dB] [mW] ×10−16 J ×10−12 J
[129] 0.5 11 62 1.7 980 75 64 DT, fb+ff 4 1

[130] 0.18 16 70 0.2 13 2.4 48 DT, fb 3 1
[131] 0.35 20 105 55 0.87 9.5 128 DT, ff 5 4
[132] 0.35 20 78 5.6 44 22 256 DT, fb 2 1
[133] 0.5 20 100.3 66 3.1 19 76 DT, fb 5 5
[42] 0.09 20 81 0.13 0.52 0.35 100 DT, fb 3 1
[134] 0.5 24 98 90 5.9 29 128 DT, fb+feedin 2 4
[36] 0.5 25 70 0.135 5.4 1.0 48 CT, fb 3 1
[50] 0.13 25 71 0.25 7.9 1.7 48 CT, fb 3 1
[135] 0.35 25 85 0.95 1.2 1.3 100 DT, fb 3 1
[132] 0.35 50 74 5.6 45 14 102 DT, fb 2 1
[123] 0.13 100 74.1 1.28 5 1.5 500 DT 2 1
[136] 0.13 100 81 2.4 1.9 1.3 195 DT, fb 2 1
[137] 0.18 100 79 5 6.3 3.4 130 CT, fb 3 1
[38] 0.35 100 82 1.8 1.1 0.87 65 CT, ff 4 1
[138] 0.5 135 66 4 74 9.1 48 DT, fb 2 1
[139] 0.4BiC. 180 82 5 1.8 1.3 36 DT, fb 2-2 1-1
[31] 0.18 200 64 1.75 35 3.4 50 CT, fb 2 4
[140] 0.35 200 42.3 12 35000 280 107 DT, fb, BP 2 1
[141] 0.35 200 61 76 3020 210 107 DT 1
[142] 0.18 200 81 30 12 8.2 58 DT, fb 2 6
OVERVIEW OF PUBLISHED ADCs
µm [kHz] [dB] [mW] ×10−16 J ×10−12 J
[143] 0.25 200 72 11.5 36 8.8 192 DT, ff 3 1

[2] 0.25 200 82 8 2.5 1.9 52 CT 5 1
[144] 0.8 250 86 30 3 3.7 96 DT, fb 5 1
[102] 0.25 250 77.4 77 56 25 20 DT, fb, BP 2-2 3-3
[145] 0.6 250 94 210 3.3 10 64 DT, ff 5 1
[146] 0.25 270 78.4 9.2 4.9 2.5 48 CT, ff 4 1
[147] 0.35 270 78 24 14 6.8 148 DT, fb, BP 2-2 1-1
[148] 0.35 270 80 56 21 13 148 DT, fb, BP 4 1
[149] 0.35BiC. 270 81 50 15 10 48 DT+CT, fb, BP 6 3
[84] 0.18 271 90 8.1 0.3 0.58 48 CT, ff 5 1
[150] 0.18 276 78 15 8.6 4.2 96 DT, fb 2 3

[125] 0.35BiC. 500 77 12 4.8 2.1 32 DT, fb 2 4
[142] 0.18 625 77 30 10 4.1 18 DT, fb 2 6
[33] 0.18 1000 64 2.2 8.8 0.85 24 CT, fb 2 4
[123] 0.13 1000 58 1.28 20 0.99 50 DT 2 1
[32] 0.18 1000 75.5 4.4 1.2 0.44 32 CT, ff 5 1
[31] 0.18 1000 51 1.75 140 3.0 10 CT, fb, BP 2 4
[39] 0.65 1000 56.7 21.8 470 20 50 CT, fb+feedin, BP 2 1
[151] 0.18 1000 59 2 25 1.4 32 CT, ff 4 1
[28] 0.18 1000 77.3 6 1.1 0.5 141 CT, ff 3 3.2
[152] 0.18 1000 88 230 3.6 5.6 29 DT, fb 2-1-1 1-1-2
[124] 0.5 1100 84 62 2.2 2.2 16 CT, ff+feedin 3 5
195
196
µm [kHz] [dB] [mW] ×10−16 J ×10−12 J
[43] 0.13 1100 78 7 1 0.49 192 DT, fb+feedin 2 3

[84] 0.18 1230 83 8.7 0.35 0.31 31 CT, ff 5 1
[147] 0.35 1250 75 37 9.4 3.2 32 DT, fb, BP 4-4 1-1
[153] 0.25 1250 80 100 8 4.9 2 DT, ff 4 4
[154] 0.5 1250 90 105 0.84 1.6 8 DT, fb 2-1-1 4-4-4
[49] 0.65 1250 89 295 3 5.1 24 DT, fb 2-1-1 1-1-4
[147] 0.35 1762 69 37 26 4.6 22 DT, fb, BP 4-4 1-1
[142] 0.18 1920 70 50 26 5.0 6 DT, fb 2 6
[155] 0.13 1920 50.9 1.5 64 1.4 16 CT, fb 2 2.3
[136] 0.13 1920 64 4.3 8.9 0.86 10 DT, fb 2-1 1-2.3
[156] 0.12 2000 60 3 15 0.92 26 CT, ff 3 3
[123] 0.13 2000 45.2 1.28 190 2.2 25 DT 2 1
[153] 0.25 2000 74 105 21 6.4 12 DT, ff 4 4
[157] 0.18 2000 82 150 4.7 3.6 8 DT, ff+fb, BP 5 4
[37] 0.18 2000 68 3.3 2.6 0.4 38 CT, ff 4 2
[158] 0.5 2000 87 150 1.5 2.0 16 DT,fb 2-2-1 5-3-3
[159] 0.25 2000 70 110 55 11 16 DT 2-2 1-1
[160] 0.18 2500 69 150 76 13 24 DT, BP 2-2-2 3-5-5
[161] 0.6 3100 56 16 130 5.0 64 CT, ff+fb 5 1
[147] 0.35 3840 48 38 1570 24 10 DT, fb, BP 4-4 1-1
[84] 0.18 3840 72 9.5 1.6 0.38 20 CT, ff 5 1
[148] 0.35 3840 42 56 9200 71 10 DT, fb, BP 4 1
µm [kHz] [dB] [mW] ×10−16 J ×10−12 J
[143] 0.25 3840 52 13.5 220 5.4 24 DT, ff 3 1

[162] 0.65 6250 67 380 120 17 8 DT, fb 3 4
[163] 0.18 12000 72 200 11 2.6 8 DT, fb 5 4
[34] 0.13 15000 63.7 70 20 1.9 10 CT, ff+fb 4 4
[85] 0.18 20000 56 122 150 5.9 8 CT, ff 2-2 4-4
[6] 0.13 20000 50 80 400 7.7 4 DT, fb 2-2 1.5-4
[6] 0.13 40000 50 106 270 5.1 4 DT, fb 2-2 1.5-4
197
Appendix B
Power/performance relation of
analog circuits
First order relations between the current consumption on one side and noise, distortion
and signal level on the other side are derived for the topologies with differential pair in
figure 4.2. It is assumed that the transconductance of the input transistors constitute the
dominant noise and distortion source.
B.1 Simple differential pair

Some well-known equations for noise and distortion of the simple differential pair of
figure 4.2.a can be calculated in a straight-forward way.
gm v̂IN2 /2
DR = (B.1)
(8kTγ BW)

3 v̂IN 2
IM3 = (B.2)
32 vGT
Notice v̂IN refers to the amplitude of the input signal. The parameter γ is the noise-excess
factor. In present CMOS technologies 4kTγ equals 2.1 × 10−20 J .
From the above equations a dependence of the quiescent current on the noise and
distortion metrics can be derived. It should be mentioned that often the input signals to
the conditioning channel are very small and distortion is not a problem for the first stages.
In that case, the overdrive voltage vGT can be chosen close to weak inversion (yielding the
highest transconductance for the current spent) instead of the value as calculated in (B.2).
The current consumption then is independent of the linearity requirement. In order to
generalize the results, a parameter α has been introduced: α becomes zero for relaxed
distortion demands, α equals one in case vGT is set by equations B.2:
α
DR BW v̂IN
I ∼ √ (α = 0, 1) (B.3)
v̂IN2 IM3
199
200 POWER OF ANALOG CIRCUITS
B.2 Differential pair in a global feed-back configuration

For stages further on in the channel linearity can be a problem and the vGT value required
from (B.2) may become too high. Instead, negative feed-back can be applied to linearize
the circuit. Moreover, the circuit becomes less sensitive to distortion of the output im-
pedance. The penalty of feed-back is in the higher bandwidth requirement. Referring to
figure 4.2.c, it is assumed that the closed loop gain is large such that Rin << Rfb . Within
the signal bandwidth the open loop gain |A(s)| on its turn is assumed to be much larger
than the closed loop gain R f b /Rin , hence: gm Rin >> 1. Therefore, Rin constitutes the
dominant noise impedance:
v̂IN2 /2
DR = (B.4)
8kTRin BW
Distortion is determined by the error signal v̂ at the virtual ground node compared to the
vGT of the input transistor. It can be calculated that:
Rfb
v̂ = v̂IN (B.5)
A(s)Rin
A third-order intermodulation component is generated by the non-linearity of the differ-
ential pair. It can be represented by an equivalent signal source of value:
3 3
2
v̂ (B.6)
32vGT
at the virtual ground node with [164]. The input-referred equivalent distortion source
is larger by a factor 1 + Rin /Rfb which is close to 1. It is compared to the input signal
yielding the IM3 distortion:
3
3 v̂IN 2 Rfb
IM3 = (B.7)
32 vGT A(s)Rin
This equation can be simplified in case of an operational transconductance amplifier.
Within the bandwidth of interest A(s) equals gm R f b , hence:
3
3 v̂IN 2 1
IM3 = (B.8)
32 vGT gm Rin
If the input transistors are biased near weak-inversion :
• the linear input range of the differential pair decreases
• the intermodulation component of eq. B.6 reduces because A(s) increases
The latter effect is dominant because of the cubic relation. Hence distortion decreases.
From equation B.4 the maximum input resistance Rin can be calculated. This value
is filled out in equation B.7. As the open loop gain A(s) is proportional to the quiescent
current, the following power/performance relation can be derived:
α
DR BW v̂IN
I ∼ √ (α = 2/3) (B.9)
v̂IN2 I M3
B.3. DEGENERATED DIFFERENTIAL PAIR 201
The α-parameter has been used in view of comparison with the results for the simple
differential pair.
B.3 Degenerated differential pair

An alternative solution in case of challenging linearity requirements is in applying de-
generation to the differential pair (figure 4.2.c). In fact, degeneration is a case of local
feed-back and the same power/performance relations as expressed by equation (B.9) can
be calculated.
Appendix C
Power/performance relation of
digital filters
In section 4.5.2 the power consumption of a digital filter is expressed as a function of a.o.
the sample rate and the capacitance. Here, a relation between the capacitance and the filter
specifications is derived. This is done per clock domain, i.e. per decimation stage. Next,
the filter specifications are written in terms of the parameters of the conditioning channel,
i.e. ENOB, mfs and f s . Combining these relations, an upper limit on the power/per-
formance relation of decimation filters results. The analysis is conducted assuming the
direct-form FIR implementation of fig. C.1 for all decimation sections. Obviously, this
implementation cannot be used for mapping the CIC filters. This is commented after-
wards.
C.1 Analysis of the filter topology

The implementation of figure C.1 consists of a shift register operating at the high sam-
ple rate and a computational part operating at the decimated frequency. It is assumed
that the filter coefficients are in a Canonic Signed Digit (CSD) format. In that case, the
multiplication can be rewritten in terms of additions only1 . Furthermore, the following
nomenclature is used for the filter parameters:
Ni = number of taps of decimation stage i

Wi = word-length at the input of stage i
m i f s = sample rate at the input of stage i
(m 1 equals m)
pi = average number of additions per tap of stage i
(assuming every computation can be written as an addition)
1 In the CSD format a number is expressed as a sum of powers of 2. This allows nested multiplication:
partial products are calculated using simple bit shifts and are summed afterwards. Hence, the complexity of a
multiplication is then reduced to that of adding a number of partial products.
203
204 POWER OF DIGITAL FILTERS
Figure C.1: Direct-form FIR implementation
The current consumption in the shift register of stage i can then be expressed as:
Ii,shi f t ∼ m i f s Ni Wi (C.1)
The current consumption of the computational part obeys:
Ii,comp ∼ m i−1 f s Ni Wi pi (C.2)
C.2 Calculation of filter parameters

The filter parameters Ni , pi and Wi are expressed in terms of the parameters of the con-
ditioning channel.
Ni : The number of taps Ni can be estimated using the empirical formula by Kaiser:

−10 log δ p δs − 13
Ni = 14.6 f
(C.3)
i
m i fs
f i is the transition band of the i th decimation stage. The pass-band ripple δ p was as-
sumed constant and it approximates 1 anyhow. Hence, the Kaiser formula indicates that
the number of taps is proportional to the required stop-band suppression (expressed in
dB) divided by the relative transition band of the filter stage. The required stop-band
suppression (expressed in dB) is proportional to the E N O B 2 , hence:
ENOB
Ni ∼ fi
(C.4)
m i fs
2 As the resolution requirements increase the quantization noise must be suppressed deeper.
C.3. CALCULATION OF POWER CONSUMPTION 205
pi : The number of partial additions pi (per filter tap) depends on the accuracy of
the coefficient. As this accuracy decreases the frequency response of the filter degrades
and more quantization noise folds back. A rule of thumb for the number of CSD dig-
its per coefficient -and thus for the number of additions of partial products- is given in
reference [165]: one CSD per 20dB of stop-band suppression should be counted, hence:
pi ∼ ENOB (C.5)
Wi : The word-length Wi equals 1 in the input stage (considering only single-bit
conversion). In the last stage it approximates ENOB. In the intermediate stages it has
an in-between value.
C.3 Calculation of power consumption

Relations C.1 and C.2 can now be filled out for the various decimation stages.
First decimation stage: The contributions of the shift register and the computational
part are added. N1 is calculated from eq. C.4 with a relative transition band of almost 1/2.
Wi equals 1 for single-bit A/D conversion and p1 follows from eq. C.5.

I1 ∼ ENOB mfs + ENOB2 mfs ≤ O ENOB2 mfs (C.6)
for large ENOB.

Last decimation stage: The relative transition band is set by the application (as dis-
cussed on page 48). The word-length of the data about equals the required ENOB and the
number of additions is again determined by the stop-band suppression.
 
ENOB2 ENOB3 ENOB 3
Ilast ∼ f 2fs + f fs ≤ O  f fs  (C.7)
2fs 2fs 2fs
Notice ENOB appears in a cubic relation: the number of filter taps, the word-length of
the coefficients and of the data are all proportional to ENOB. Hence, the last decimation
stage is normally the largest in terms of die area. Also, it must achieve the smallest relative
transition band of all sections. This strongly affects the number of filter taps, the area and
the power consumption. Although, this stage is sampled at an m times lower frequency
than the first stage its power consumption can be as -or even more- important due to its
complexity.
Intermediate stages: As the sample rate gets decimated the complexity (i.e. number
of taps, data word-length and coefficient word-length) of the filter stage increases. Hence,
the current consumption of the intermediate stages is expected to follow a relation in
between that of the first (eq. C.6) and the last (eq. C.7) stage.
Overall current consumption: The contributions of the various stages should be

summed in order to calculate the overall current consumption. The number of decimation
206 POWER OF DIGITAL FILTERS
stages is a logarithmic function of the over-sample ratio m. The base of the logarithm
depends on the decimation factors used; e.g. in case all stages decimate by a factor of 2
about ln2 m stages are needed.
In case the last decimation stage dominates the power consumption, a cubic depen-
dence on ENOB is valid while the over-sample ratio is of less importance. In case the first
decimation stage dominates the input frequency m f s is of utmost importance while the
dependence on ENOB is only quadratic. An upper-limit on the dependence of the overall
current consumption on the parameters of the conditioning channel is expressed by the
following relation:  
ENOB3
Iworst−case ∼ O  f
mfs ln m (C.8)
2fs
Of course, this limit combines a worst-case combination of dependencies and is pes-

simistic. Also, it only gives relative results. Hence, it is only intended for extrapolation of
the power consumption of a specific architecture to modified performance specifications.
C.4 Extension to other implementations

The power/performance analysis has been conducted for the direct-form FIR implemen-
tation of figure C.1. It can easily be understood that transposed-form, poly-phase imple-
mentations, etc. follow the same relation with a different proportionality parameter. For
the implementation of a CIC filter a similar analysis can be conducted. The result is the
same as that of eq. C.8 except for the dependence on ENOB. Here, ENOB only occurs
in a quadratic relation because CIC filters do not use coefficients (the data is integrated
instead).
Appendix D
Third-order distortion in analog

circuits and ADCs
Based on the models in fig. D.1.a and b an expression for the third-order distortion in
analog circuits and in a ADC is derived. The model in fig. D.1.a represents an analog
circuit with feedback. The model in fig. D.1.b represents a ADC: a bitstream signal
is fed back. The output of the quantizer can be scaled such that the linearized DAC gain
equals 1. It is shown that the topology of fig. D.1.b results in a two times larger third-order
distortion, even if the same input stage (thus the same non-linearity) is used in both cases.
Assuming differential operation, the input transconductor performs the following V-to-I
conversion:
i ≈ h 1 v + h 3 v3 (D.1)
where v equals the difference between output and input (assuming a gain of 1 in the
feedback). h 1 and h 3 are the transfer coefficients for the fundamental and for the 3r d -order
distortion component respectively. The higher-order distortion components are neglected.
Furthermore, a sine-wave input is assumed and the following relations are used:
v I N = sin x (D.2)
v 2I N = 1/2 − 1/2 cos 2x (D.3)
v 3I N = 3/4 sin x − 1/4 sin 3x (D.4)
In the following, the amplitude of the 3r d -order distortion component of the input stage is
calculated for the model of fig. D.1.a and for that of fig. D.1.b. Assuming equal loop gain
in both cases, the transfer function from this node to the output node is the same. Hence,
it is sufficient to calculate the h 3 -coefficient of the input stage and compare the result for
model a and b.
Analog circuit (fig. D.1.a):

207
208 THIRD-ORDER DISTORTION IN ANALOG CIRCUITS
Figure D.1: Functional model of a non-linear, analog circuit with feedback (a)
and of a non-linear ADC (b)
Assuming:
v OU T = (1 + δ)v I N (D.5)
with δ a small number. Using eq. D.4, the last term in eq. D.1 is expanded:
h3 3 3
h 3 v3 = δ vI N (D.6)
4
Hence, the amplitude of 3r d -order distortion component of the input stage equals:
h3
| | (D.7)
4
ADC (fig. D.1.b):

Again, the last term in eq. D.1 is expanded:
h3 v3 = h3 v3I N − 3h3 v2I N v OU T + 3h3 v I N v2OU T − h3 v3OU T (D.8)
Since v OU T is a bitstream encoded signal, the last two terms do not contribute any third-
harmonic distortion [40]:
• the square of a bitstream signal yields a constant
• consequently, the 3r d power of a bitstream signal yields the original bitstream
Hence, the 3r d -order distortion component of the input stage is calculated from the first
two terms of eq. D.8. The low-frequency content of the output bitstream equals the input
sine wave (except for the error δ), i.e. eq. D.5 is valid for low-frequency inputs. Fur-
thermore, relations eq. D.2 upto eq. D.4 are substituted, yielding the amplitude of the
3r d -order distortion component:
h3 3h3 h3
| − (1 + δ)| ≈ | | (D.9)
4 4 2
if δ is small. Note that this amplitude is twice larger than that of eq. D.7.
THIRD-ORDER DISTORTION IN ANALOG CIRCUITS 209
In the models of fig. D.1.a and b the same non-linearity is assumed for the input
stage of the analog circuit and for the ADC. The loop gain and, therefore, the error
δ on the output signal are assumed to be the same as well. Still, the 3r d -order distortion
component is a factor two higher for the ADC. This is due to the fact that a bitstream
signal is fed back instead of a sine wave.
Appendix E
Power consumption in a data

interface
The power consumption in an analog and a digital inter-die interface is studied. The
physical interface between the ICs may consist of bond-pads, bond wires, PCB tracks,
etc. It should be modeled by an RLC-network with a certain filtering -or even resonating-
transfer. The influence of the inductance is disregarded here. It is assumed that the reso-
nance occurs at frequencies way beyond the band of interest and/or that the resonance is
properly damped.
Furthermore, only the power consumed at the transmit side of the link is considered.
The power at the receiver side is much smaller because of an easier load (i.e. a smaller
capacitance and a higher resistance).
In all cases a differential interface is assumed.
E.1 Analog data interface

A model of the transmit side of the analog link is depicted in figure E.1. Within the
bandwidth of this link:
1
(E.1)
2πRL CL
The peak-to-peak voltage swing V pp equals:
Vpp = I RL (E.2)
It is determined by the quiescent current I through the load resistor R L . Noise is con-
tributed both by the transconductance of the transmit stage and by the resistive load:

v2 n,eq = 4kT BW γ gm R2L + RL (E.3)
211
212 POWER CONSUMPTION IN A DATA INTERFACE
Figure E.2: Transmit model for full-swing (a) and low-swing (b) digital inter-
face [89]
Assuming a gain gm R L larger than 1 the contribution of the transconductance is dominant,

hence:
I | vGT |
DR = (E.4)
16kTγ BW
The dynamic range D R can be expressed as a function of the number of bits:
3 2N
DR = 2 (E.5)
2
From the above equations the current consumption of the analog transmitter can be calcu-
lated. Normally, a substantial margin is foreseen on this value to account for distortion and
interference injected in the link. Hence, the result is given in terms of a proportionality
relation:
I ∼ 22N BW (E.6)
E.2 Digital data interface

This analysis is partly taken from [89] and is recapitulated and supplemented here. A full-
swing and a low-swing (or slew-rate controlled) interface are distinguished (figure E.2.a
and b respectively). A low-swing digital interface generates less interference.
E.2. DIGITAL DATA INTERFACE 213
It should be noticed that a digital communication link requires synchronization at the

receiver side. This can be done by either transmitting the clock or recovering it from the
data. The associated power is not considered here.
Furthermore, the average switching frequency f sw is introduced: it is defined as the
inverse of the average number of 0−1 transitions in the bit stream. Since f sw is data-
dependent it results that the power consumption in a digital link is data dependent as well.
Full-swing interface A full-swing interface normally uses an inverter type of trans-

mitter as depicted in figure E.2.a. The bandwidth of the link is limited by the slew-rate of
the inverters for the specified load. It can be calculated that the average current consump-
tion of the pseudo-differential transmitter equals:
Ifull−swing = CL Vdd N fsw (E.7)
The parameter N indicates the number of bits that must be transmitted either via some
parallel links or via the same link after serializing the data.
Low-swing interface For the low-swing interface a termination resistor RT is in-

troduced. Its value is either standardized or it follows from the following bandwidth
requirements:
1
f sw for a parallel interface (E.8)
2π RT C L
1
Nfsw for a serialized interface (E.9)
2π RT C L
Part of the current flows into RT , the other part is charging and discharging the load
capacitor C L . The current consumption in the interface is given by:
C L f sw
Ilow−swing = (RT−1 + )V pp N for a parallel interface (E.10)
2
C L N f sw
Ilow−swing = (RT−1 + )V pp for a serialized interface (E.11)
2
In both cases, the capacitive part of the current consumption (i.e. the dynamic current) can
be neglected compared to the dissipation in the termination resistor (i.e. the static current)
because of the bandwidth requirement (eq. E.9):
Ilow−swing ∼
= RT−1 V pp N for a parallel interface (E.12)
Ilow−swing ∼
= RT−1 V pp for a serialized interface (E.13)
Serializing the parallel bitstreams is favorable for power consumption since the available
bandwidth is used more effectively. Comparing the result to that for the full-swing in-
terface (eq. E.7) it can be concluded that, in case of a low-resolution link, the full-swing
214 POWER CONSUMPTION IN A DATA INTERFACE
interface is lower power because it does not use any static current. For higher data rates
or a large load capacitance a low-swing interface should be favored.
In case the value of RT is fixed by a standard or application the current consumption
in the low-swing link is independent of f sw as long as the bandwidth requirements is
fulfilled. It scales with the number of parallel bits.
In case RT is not standardized, the link can be designed for a minimum current
consumption by choosing RT as high as possible. The upper limit on RT is given by
eq. E.9. When scaling RT to this limit, the dynamic and the static contribution in eq. E.13
become equally important. Eq. E.11 can be simplified such that only the dependence on
the number of bits N and on the average switching frequency f sw remains:
Ilow−swing ∼ N fsw (E.14)
This relation holds both for the parallel and the serialized link. Also, the scaling with N
and f sw is the same as for the full-swing interface (eq. E.7) with a capacitive load only.
References
[1] M. Pelgrom and M. Vertregt, “CMOS technology for mixed-signal IC’s,” Solid-
State Electron., vol. 41, no. 7, pp. 967–974, 1997.
[2] E. van der Zwan, K. Philips, and C. Bastiaansen, “A 10.7MHz IF-to-baseband

A/D conversion system for AM/FM radio receivers,” IEEE J. Solid-State Cir-
cuits, vol. 35, no. 12, pp. 1810–1819, Dec. 2000.
[3] W. Sansen, M. Steyaert, V. Peluso, and E. Peeters, “Toward sub 1V analog Inte-
grated Circuits in submicron standard CMOS technologies,” in ISSCC Digest of
Tech. Papers, vol. 41, Feb. 1998, pp. 186–187, 435.
[4] A. Abidi, G. Pottie, and W. Kaiser, “Power-conscious design of wireless circuits

and systems,” in Proc. IEEE, vol. 88, Oct. 2000, pp. 1528–1545.
[5] A. Annema, “Analog circuit performance and process scaling,” IEEE Trans. Cir-
cuits and Systems II, vol. 46, no. 6, pp. 711–715, June 1999.
[6] A. Tabatabaei, K. Onodera, M. Zargari, H. Samavati, and D. Su, “A dual channel

ADC with 40MHz aggregate signal bandwidth,” in ISSCC Digest of Tech.
Papers, vol. 46, Feb. 2003, pp. 66–67, 478.
[7] E. van der Zwan and E. C. Dijkmans, “A 0.2mW CMOS modulator for speech
coding with 80dB dynamic range,” IEEE J. Solid-State Circuits, vol. 31, no. 12,
pp. 1873–1880, Dec. 1996.
[8] A. Annema, “CMOS technology evolution and analog circuit performance,” in

Proc. 12th ProRISC Workshop, 2001, pp. 245–249.
[9] A.J.Scholten, L. Tiemeijer, R. van Langevelde, R. Havens, A. T. A. Z. van Duijn-

hoven, and V. Venezia, “Noise modeling for RF CMOS circuit simulation,” IEEE
Trans. Electron. Devices, vol. 50, no. 3, pp. 618–632, Mar. 2003.
[10] R. van Langevelde, “A compact MOSFET model for distortion analysis in analog
circuit design,” Ph.D. dissertation, Eindhoven Univ. of Techn., The Netherlands,
Nov. 1998.
215
216 REFERENCES
[11] A. Scholten, H. Tromp, L. Tiemeijer, R. van Langevelde, R. Havens, P. de Vreede,

R. Roes, P. Woerlee, A. Montree, and D. Klaassen, “Accurate thermal noise model
for deep-submicron CMOS,” in IEDM Digest of Tech. Papers, Dec. 1999, pp. 155–
158.
[12] W. Sansen, “Analog circuit design in scaled CMOS technology,” in VLSI Circuits
Digest of Tech. Papers, June 1996, pp. 8–11.
[13] Q. Huang, F. Piazza, P. Orsatti, and T. Ohguro, “The impact of scaling down to deep
submicron on CMOS RF circuits,” IEEE J. Solid-State Circuits, vol. 33, no. 7, pp.
1023–1036, July 1998.
[14] International Technology Roadmap for Semiconductors 2003 edition. ITRS.

[Online]. Available: http://public.itrs.net/Files/2003ITRS/Home2003.htm
[15] Technical specification group radio access networks; UE radio transmission and
reception (FDD), 3r d generation partnership project Std. release 99 3G TS 25.213,
v3.2.0.
[16] B. Minnis and P. Moore, “A highly digitized multimode receiver architecture for
3G mobiles,” IEEE Trans. Vehicular Technology, vol. 52, no. 3, pp. 637–653, May
2002.
[17] ——, A qualitative review of UMTS transmitter architectures. PRL Tech. Note
4061, Philips internal report, 2001.
[18] A. Caswell and I. Ozerin, Dimensioning a polar loop for a UMTS transmitter.
PRL Tech. Note 4099, Philips internal report, 2001.
[19] K. Philips and B. Minnis, Considerations on an A-to-D converter for UMTS.

Nat.Lab. Tech. Note 425, Philips internal report, 1999.
[20] B. Murmann and B. Boser, “Digitally assisted analog integrated circuits,” DSP,
vol. 2, no. 1, pp. 2040–2050, Mar. 2004.
[21] “French patent presentation,” 987283, May 22, 1948.
[22] F. de Jager, “Delta modulation, a method of PCM transmission using the one unit
code,” Philips Res. Rep., vol. 7, pp. 442–466, 1952.
[23] C. Cutler, “Transmission systems employing quantization,” U.S. Patent2,927,962,

March 8, 1960.
[24] H. Inose, Y. Yasuda, and J. Murakami, “A telemetering system by code modulation-

modulation,” IRE Trans. Space Electron. Telemetry, vol. SET-8, pp. 204–2–9,
Sept. 1962.
[25] S. Norsworthy, R. Schreier, and G. Temes, Delta-Sigma Data Converters. New

York: IEEE Press, 1996.
REFERENCES 217
[26] J. van Engelen and R. van de Plassche, Bandpass Sigma delta modulators. Kluwer
Academic Publishers, 1999.
[27] J. Cherry and W. Snelgrove, Continuous-time Delta-Sigma Modulators for high-

speed A/D conversion. Norwell, MA: Kluwer Academic Publishers, 2000.
[28] B. Putter, “ ADC with finite impulse response feedback DAC,” in ISSCC Digest
of Tech. Papers, Feb. 2004, pp. 76–77.
[29] J. Yu and F. Maloberti, “A low-power multi-bit modulator in 90nm digital

CMOS without DEM,” in ISSCC Digest of Tech. Papers, Feb. 2005, pp. 168–169,
591.
[30] R. van Veldhoven, “A tri-mode continuous-time modulator with switched-

capacitor feedback DAC for a GSM-EDGE/CDMA2000/UMTS receiver,” in
ISSCC Digest of Tech. Papers, Feb. 2002, pp. 60–61, 477.
[31] M. Kappes, H. Jensen, and T. Gloerstad, “A versatile 1.75mW CMOS continuous-

time - ADC with 75dB dynamic range for wireless applications,” in ESSCIRC
Digest of Tech. Papers, Sept. 2002, pp. 279–282.
[32] K. Philips, “A 4.4mW 76dB complex ADC for Bluetooth receivers,” in ISSCC
Digest of Tech. Papers, Feb. 2003, pp. 64–65,478.
[33] M. Kappes, “A 2.2-mW CMOS bandpass continuous-time multibit - ADC

with 68dB of dynamic range and 1MHz bandwidth for wireless applications,” IEEE
J. Solid-State Circuits, vol. 38, no. 7, pp. 1098–1104, July 2003.
[34] S. Paton, A. D. Giandomenico, L. Hernandez, A. Wiesbauer, T. Potscher, and

M. Clara, “A 70mW 300MHz CMOS continuous-time ADC with 15MHz
bandwidth and 11 bits of resolution,” IEEE J. Solid-State Circuits, vol. 39, no. 7,
pp. 1056–1063, July 2004.
[35] P. Vancorenland, P. Coppejans, W. D. Cock, and M. Steyaert, “A quadrature digital

downconverter,” in CICC Digest of Tech. Papers, 2002, pp. 235–238.
[36] F. Gerfers, M. Ortmanns, and Y. Manoli, “A 1.5V 12-bit power-efficient

continuous-time third-order modulator,” IEEE J. Solid-State Circuits, vol. 38,
no. 8, pp. 1343–1352, Aug. 2003.
[37] R. van Veldhoven, B. Minnis, and H. Hegt, “A 3.3-mW modulator for UMTS
in 0.18µm CMOS with 70-dB dynamic range in 2-MHz bandwidth,” IEEE J. Solid-
State Circuits, vol. 37, no. 12, pp. 1645–1661, Dec. 2002.
[38] L. Breems, E. van der Zwan, and J. Huijsing, “A 1.8-mW CMOS modula-
tor with integrated mixer for A/D conversion of IF signals,” IEEE J. Solid-State
Circuits, vol. 35, no. 4, pp. 468 –475, Apr. 2000.
218 REFERENCES
[39] F. Henkel, U. Langmann, A. Hanke, S. Heinen, and E. Wagner, “A 1-MHz-

bandwidth second-order continuous-time quadrature bandpass sigma-delta mod-
ulator for low-IF radio receivers,” IEEE J. Solid-State Circuits, vol. 37, no. 12, pp.
1628–1635, Dec. 2002.
[40] L. Breems and J. Huijsing, Continuous-Time Sigma-Delta Modulation for A/D

Conversion in Radio Receivers. Kluwer Academic Publishers, 2001.
[41] H. van der Ploeg, “Calibration techniques in two-step A/D converters,” Ph.D. dis-
sertation, Univ. Twente, The Netherlands, Mar. 2005.
[42] L. Yao, M. Steyaert, and W. Sansen, “A 1V 88dB 20kHz modulator in 90nm

CMOS,” in ISSCC Digest of Tech. Papers, Feb. 2004, pp. 80–81, 514.
[43] R. Gaggl, M. Inversi, and A. Wiesbauer, “A power optimized 14-bit SC

modulator for ADSL CO applications,” in ISSCC Digest of Tech. Papers, Feb.
2004, pp. 82–83, 514.
[44] W. Gao, O. Shoaei, and W. Snelgrove, “Excess loop delay effects in continuous-
time delta-sigma modulators and the compensation solution,” in Proc. Int. Symp.
Circuits Syst., vol. 1, 1997, pp. 65–68.
[45] J. Cherry and W. Snelgrove, “Excess loop delay in continuous-time delta-sigma

modulators,” IEEE Trans. Circuits and Systems II, vol. 46, no. 4, pp. 376–389,
Apr. 1999.
[46] G. Hurkx, P. Agarwal, R. Dekker, E. van der Heijden, and H. Veenstra, “RF figures-
of-merit for process optimization,” IEEE Trans. Electron. Devices, vol. 51, no. 12,
pp. 2121 – 2128, Dec. 2004.
[47] T. Burger and Q. Huang, “A 13.5mW, 185MSample/s -modulator for

UMTS/GSM dual-standard IF reception,” in ISSCC Digest of Tech. Papers, Feb.
2002, pp. 44–45, 427.
[48] H. Tao and J. M. Khoury, “A 100MHz IF, 400MSample/s CMOS direct-conversion

bandpass modulator,” in ISSCC Digest of Tech. Papers, Feb. 1999, pp. 60–61,
445.
[49] Y. Geerts, M. Steyaert, and W. Sansen, “A 2.5MSample/s multi-bit CMOS

ADC with 95dB SNR,” IEEE J. Solid-State Circuits, vol. 35, no. 12, pp. 1829–
1839, Dec. 2000.
[50] M. Ortmanns and F. Gerfers, “A continuous-time modulator with reduced jitter

sensitivity,” in ESSCIRC Digest of Tech. Papers, Sept. 2002, pp. 287–290.
[51] S. Luschas and H. Lee, “High-speed modulators with reduced timing jitter
sensitivity,” IEEE Trans. Circuits and Systems II, vol. 49, no. 11, pp. 712–720,
2002.
REFERENCES 219
[52] R. Adams and K. Nguyen, “A 113dB SNR oversampling DAC with segmented
noise-shaping scrambling,” in ISSCC Digest of Tech. Papers, Feb. 1998, pp. 1871
– 1878.
[53] P. Allen, “Slew induced distortion in operational amplifiers,” IEEE J. Solid-State
Circuits, vol. 12, no. 1, pp. 39–44, Feb. 1977.
[54] W. Sansen, H. Qiuting, and K. Halonen, “Transient analysis of charge transfer in
SC filters - gain error and distortion,” IEEE J. Solid-State Circuits, vol. 22, no. 2,
pp. 268–276, Apr. 1987.
[55] E. Vittoz, “Future of analog in the VLSI environment,” IEEE Int. Symp. on Circuits
and Systems, vol. 2, pp. 1372 –1375, May 1990.
[56] J. Meindl and J. Davis, “The fundamental limit on binary switching energy for
terascale integration (TSI),” IEEE J. Solid-State Circuits, vol. 35, no. 10, pp. 1515–
1516, Oct. 2000.
[57] E. Dijkstra, O. Nys, and E. Blumenkrantz, Low power Oversampled A/D convert-
ers. Kluwer Academic Publishers, 1997, pp. 89–103.
[58] A.-J. Annema, B. Nauta, R. van Langevelde, and H. Tuinhout, “Analog circuits in
ultra-deep-submicron CMOS,” IEEE J. Solid-State Circuits, vol. 40, no. 1, pp. 132
– 143, Jan. 2005.
[59] M. Steyaert, V. Peluso, J. Bastos, P. Kinget, and W.Sansen, “Custom analog low
power design: the problem of low voltage and mismatch,” in CICC Digest of Tech.
Papers, May 1997, pp. 285 – 292.
[60] D. Robertson, “Specifications and Figures of Merit for mixed-signal circuits,”
ISSCC Tutorial, 2002.
[61] B. Murmann, “Limits on ADC power dissipation,” in Workshop on Advances in
Analog Circuit Design (AACD), Apr. 2005, accepted for publication.
[62] J. Voorman, “Transconductance amplifier,” U.S. Patent4,723,110, Feb. 2, 1988.
[63] J. Voorman, W. Bruls, and P. Barth, “Bipolar integration of analog gyrator and
Laguerre type filters,” in Proc. ECCTD, Sept. 1983, pp. 108–110.
[64] B. Nauta, “A CMOS transconductance-C filter for very high frequencies,” IEEE J.
Solid-State Circuits, vol. 27, no. 2, pp. 142–153, Feb. 1992.
[65] B. Gilbert, “A high-performance monolithic multiplier using active feedback,”
IEEE J. Solid-State Circuits, vol. 9, no. 6, pp. 364–373, Dec. 1974.
[66] R. Crochiere and L. Rabiner, Multirate digital signal processing. Englewood
Cliff, NJ: Prentice Hall, 1983.
[67] ——, “Interpolation and decimation of digital signals: a tutorial review,” Proc.
IEEE, vol. 69, no. 3, pp. 300–331, Mar. 1981.
220 REFERENCES
[68] D. Goodman and M. Carey, “Nine digital filters for decimation and interpolation,”
IEEE Trans. Acoust., Speech, Signal Processing, vol. 25, no. 2, pp. 121–126, Apr.
1977.
[69] J. Candy, “Decimation for sigma-delta modulation,” IEEE Trans. Communications,
vol. 34, no. 1, pp. 72–76, Jan. 1986.
[70] E. Hogenauer, “An economical class of digital filters for decimation and interpola-
tion,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 29, no. 2, pp. 155–162,
Apr. 1981.
[71] D. Tufts, D. Rorabacher, and W. Mosier, “Designing simple, effective digital fil-
ters,” IEEE Trans. Audio Electroacoust., vol. 18, pp. 142–158, June 1970.
[72] H. Veendrick, Deep-submicron CMOS ICs. Kluwer Bedrijfsinformatie, 1998.
[73] D. Birru, “Reduced-sample-rate sigma-delta modulation,” Ph.D. dissertation, Delft
Univ. of Techn., The Netherlands, June 1998.
[74] R. Schreier, “An empirical study of high-order single-bit delta-sigma modulators,”
IEEE Trans. Circuits and Systems II, vol. 40, no. 8, pp. 461–467, Aug. 1993.
[75] B. Boser and B. Wooley, “The design of sigma-delta modulation analog-to-digital
converters,” IEEE J. Solid-State Circuits, vol. 23, pp. 1298–1308, Dec. 1981.
[76] R. Adams, “Design and implementation of an audio 18-bit analog-to-digital con-
verter using oversampling techniques,” in J. Audio Eng. Soc., vol. 35, Mar. 1986,
pp. 153–166.
[77] R. Adams, P. Ferguson, A. Ganesan, A. V. S. Vincelette, and R. Libert, “Theory and
practical implementation of a fifth-order sigma-delta A/D converter,” in J. Audio
Eng. Soc., vol. 39, July 1991, pp. 515–528.
[78] J. Candy and O. Benjamin, “The structure of quantization noise from sigma-delta
modulation,” IEEE Trans. Communications, vol. 29, no. 9, pp. 1316–1323, Sept.
1981.
[79] R. Gray, “Quantization noise spectra,” IEEE Trans. Inform. Theory, vol. 36, no. 6,
pp. 1220–1244, Nov. 1990.
[80] D. Reefman and E. Janssen, “DC analysis of high order Sigma Delta modulators,”
in Proc. AES 113th Convention, paper 5693, Sept. 2002.
[81] E. Roza, “Analog-to-digital conversion via duty-cycle modulation,” IEEE Trans.
Circuits and Systems II, vol. 44, pp. 907–914, 1997.
[82] D. Reefman and E. Janssen, Description of limit cycles in Sigma Delta Modulators.
PRLE Tech. Note 201028, Philips internal report, 2004.
[83] V. Friedman, “The structure of limit cycles in sigma-delta modulation,” IEEE
Trans. Communications, vol. 36, no. 8, pp. 972–979, Aug. 1998.
REFERENCES 221
[84] R. van Veldhoven, “A triple-mode continuous-time modulator with switched-

capacitor feedback DAC for a GSM-EDGE/CDMA200/UMTS receiver,” IEEE J.
Solid-State Circuits, vol. 38, no. 12, pp. 2069–2076, Dec. 2003.
[85] L. Breems, R. Rutten, and G. Wetzker, “A cascaded continuous-time modulator
with 67-dB dynamic range in 10-MHz bandwidth,” IEEE J. Solid-State Circuits,
vol. 39, no. 12, pp. 2152– 2160, Dec. 2004.
[86] DigRF: the digital interface standard. Digital Interface Working Group. [Online].
Available: www.digrf.com
[87] [Online]. Available: www.mipi.org
[88] JC−61 White Paper. JEDEC. [Online]. Available:
www.jedec.org/Home/jc−61−Info/TechForum579 − J C − 61−Presentation.pdf
[89] G. den Besten, “Embedded low-cost 1.2Gb/s inter-IC serial data link in 0.35µm
CMOS,” in ISSCC Digest of Tech. Papers, vol. 43, Feb. 2000, pp. 250–251.
[90] J. W. Whikehart, “DSP-based radio with IF processing,” presented at the SAE
World Cong., Mar. 2000.
[91] K. Kianush and C. Vaucher, “A global car radio IC with inaudible signal quality
checks,” in ISSCC Digest of Tech. Papers, Feb. 1998, pp. 130–131.
[92] E. Cherry and D. Hooper, Amplifying devices and low-pass amplifier design. New
York: John Wiley and Sons, 1968.
[93] K. Philips and E. Dijkmans, “A variable gain IF amplifier with −67d Bc I M 3 -
distortion at 1.4V pp output in 0.25 µm CMOS,” in VLSI Circuits Digest of Tech.
Papers, June 2001, pp. 81–82.
[94] J. Rijns, “CMOS low-distortion high-frequency variable-gain amplifier,” IEEE J.
Solid-State Circuits, vol. 31, pp. 1029–1034, July 1996.
[95] C.-C. Hsu and J.-T. Wu, “A highly linear 125MHz CMOS switched-resistor
programmable-gain amplifier,” IEEE J. Solid-State Circuits, vol. 38, no. 10, pp.
1663–1670, Oct. 2003.
[96] L. Vogt, D. Brookshire, S. Lottholz, and G. Zwiehoff, “A two-chip digital car ra-
dio,” in ISSCC Digest of Tech. Papers, Feb. 1996, pp. 350–351.
[97] S. Jantzi, W. Snelgrove, and P. Ferguson, “A fourth-order band-pass sigma-delta
modulator,” IEEE J. Solid-State Circuits, vol. 28, pp. 282–291, Mar. 1993.
[98] F. W. Singor and W. M. Snelgrove, “Switched-capacitor bandpass delta sigma A/D
modulation at 10.7MHz,” IEEE J. Solid-State Circuits, vol. 30, pp. 184–192, Mar.
1995.
[99] J. Park, E. Joe, M.-J. Choe, and B.-S. Song, “A 5MHz IF digital FM demodulator,”
IEEE J. Solid-State Circuits, vol. 34, pp. 3–11, Jan. 1999.
222 REFERENCES
[100] J. van Engelen, R. van de Plassche, E. Stikvoort, and A. Venes, “A sixth-order

continuous-time bandpass sigma-delta modulator for digital radio IF,” IEEE J.
Solid-State Circuits, vol. 34, pp. 1753–1764, Dec. 1999.
[101] A. Ong and B. Wooley, “A two-path bandpass sigma-delta modulator for digital IF
extraction at 10MHz,” IEEE J. Solid-State Circuits, vol. 32, pp. 1920–1934, Dec.
1997.
[102] T. Ueno, A. Yasuda, T. Yamaji, and T. Itakura, “A fourth-order bandpass

modulator using second-order bandpass noise-shaping dynamic element
matching.” IEEE J. Solid-State Circuits, vol. 37, no. 7, pp. 809–816, July 2002.
[103] I. Galton, W. Huff, P. Carbone, and E. Siragusa, “A delta-sigma PLL for 14-b 50-
ksample/s frequency-to-digital conversion of a 10MHz FM signal,” IEEE J. Solid-
State Circuits, vol. 33, pp. 2042–2052, Dec. 1998.
[104] J. Crols and M. Steyaert, “Low IF topologies for high-performance analog front-
ends of fully integrated receivers,” IEEE Trans. Circuits and Systems II, vol. 45,
no. 4, pp. 269–282, Mar. 1998.
[105] T. Lee, The design of CMOS radio-frequency integrated circuits. Cambridge,

U.K.: Cambridge Univ.Press, 1998.
[106] J. Volder, “The CORDIC trigonometric computing technique,” IRE Trans. Elec-
tron. Comput., vol. EC-8, pp. 330–334, Sept. 1959.
[107] Q. Sandifort, L. Breems, C. Dijkmans, and H. Schuurmans, “IF-to-digital converter

for FM/AM/IBOC radio,” in ESSCIRC Digest of Tech. Papers, Sept. 2003, pp. 707
– 710.
[108] P. Quinn, K. van Hartingsveldt, and A. van Roermund, “A 10.7-MHz CMOS SC

radio IF filter using orthogonal hardware modulation,” IEEE J. Solid-State Circuits,
vol. 35, no. 12, pp. 1865–1876, Dec. 2000.
[109] A. Nagari and G. Nicollini, “A 3v 10MHz pseudodifferential SC bandpass filter

using gain enhancement replica amplifier,” IEEE J. Solid-State Circuits, vol. 33,
pp. 626–630, Apr. 1998.
[110] J. Silva-Martinez, M. Steyaert, and W. Sansen, “A 10.7MHz 68dB SNR CMOS

continuous-time filter with on-chip automatic tuning,” in ISSCC Digest of Tech.
Papers, Feb. 1992, pp. 66–67.
[111] B.-S. Song, “A 10.7MHz switched-capacitor bandpass filter,” IEEE J. Solid-State

Circuits, vol. 24, pp. 320–324, Apr. 1989.
[112] Bluetooth standard. Bluetooth Special Interest Group. [Online]. Available:

www.bluetooth.com
REFERENCES 223
[113] S. Byun, C.-H. Park, Y. Song, S. Wang, C. Conroy, and B. Kim, “A low-power
CMOS Bluetooth RF transceiver with a digital offset canceling DLL-based GFSK
demodulator,” IEEE J. Solid-State Circuits, vol. 38, no. 10, pp. 1609–1618, Oct.
2003.
[114] B. Xia, C. Xin, W. Sheng, A. Y. Valero-Lopez, and E. Sanchez-Sinencio, “A GFSK

demodulator for low-IF Bluetooth receiver,” IEEE J. Solid-State Circuits, vol. 38,
no. 8, pp. 1397–1400, Aug. 2003.
[115] H. Darabi, S. Khorram, H.-M. Chien, M.-A. Pan, S. Wu, S. Moloudi, J. C. Leete,
J. J. Rael, M. Syed, R. Lee, B. Ibrahim, M. Rofougaran, and A. Rofougaran, “A
2.4GHz CMOS transceiver for Bluetooth,” IEEE J. Solid-State Circuits, vol. 36,
no. 12, pp. 2016–2024, Dec. 2001.
[116] W. Sheng, B. Xia, A. E. Emira, C. Xin, A. Y. Valero-Lopez, S. T. Moon, and

E. Sanchez-Sinencio, “A 3V, 0.35µm CMOS Bluetooth receiver IC,” IEEE J. Solid-
State Circuits, vol. 38, no. 1, pp. 30–42, Jan. 2003.
[117] P. Laurent, “Exact and approximate construction of digital phase modulations by

superposition of Amplitude Modulated Pulses(AMP),” IEEE Trans. Communica-
tions, vol. COM-34, no. 2, pp. 150–160, Feb. 1986.
[118] N. Filiol, N. Birkett, J. Cherry, C. Cojocaru, A. Namdar, T. Pamir, K. Sheikh,

G. Glandon, D. Payer, A. Swaminathan, R. Forbes, T. Rilley, S. M. Alinoor,
E. Macrobbie, M. Clouter, S. Pipilos, and T. Varelas, “A 22mW Bluetooth RF
transceiver with direct RF demodulation and on-chip IF filtering,” in ISSCC Digest
of Tech. Papers, Feb. 2001, pp. 202–203,447.
[119] M. Kokubo, M. Shida, T. Ishikawa, H. Sonoda, K. Yamamoto, T. Matsuura,

M. Matsuoka, T. Endo, T. Kobayashi, K. Oosaki, T. Henmi, J. Kudoh, and H. Miya-
gawa, “A 2.4GHz RF transceiver with digital channel-selection filter for Blue-
tooth,” in ISSCC Digest of Tech. Papers, Feb. 2002, pp. 94–95, 449.
[120] R. Staszewski, K. Muhammad, D. Leipold, C.-M. Hung, Y.-C. Ho, J. Wallberg,

C. Fernando, K. Maggio, R. Staszewski, T. Jung, J. Koh, S. John, I. Deng, V. Sarda,
O. Moreira-Tamayo, V. Mayega, R. Katz, O. Friedman, O. Eliezer, E. de Obaldia,
and P. Balsara, “All-digital TX frequency synthesizer and discrete-time receiver for
Bluetooth radio in 130-nm CMOS,” IEEE J. Solid-State Circuits, vol. 39, no. 12,
pp. 2278– 2291, Dec. 2004.
[121] H. Bergveld, K. van Kaam, D. Leenaerts, K. Philips, A. W. P. Vaassen, and

G. Wetzker, “A low-power highly-digitized receiver for 2.4GHz-band GFSK ap-
plications,” in RFIC Digest of Tech. Papers, June 2004, pp. 347 – 350.
[122] L. Eggermont, “Measured signal/quantising-distortion ratio of high-information

delta modulation,” Electronics Letters, vol. 11, no. 11, pp. 242–244, May 1975.
224 REFERENCES
[123] F. Chen and S. R. B. Bakkaloglu, “A 1.5V 1mA 80dB passive ADC in 0.13µm
digital CMOS process,” in ISSCC Digest of Tech. Papers, vol. 1, Feb. 2003, pp. 54
– 55, 477.
[124] S. Yan and E. Sanchez-Sinencio, “A continuous-time sigma-delta modulator with

88-dB dynamic range and 1.1-MHz signal bandwidth,” IEEE J. Solid-State Cir-
cuits, vol. 36, no. 12, pp. 75 – 86, Dec. 2001.
[125] J. Grilo, I. Galton, K. Wang, and R. Montemayor, “A 12-mW ADC delta-sigma

modulator with 80dB of dynamic range integrated in a single-chip Bluetooth trans-
ceiver,” IEEE J. Solid-State Circuits, vol. 37, no. 3, pp. 271–278, Mar. 2002.
[126] K. Doris, “High-speed D/A converters,” Ph.D. dissertation, Eindhoven Univ. of

Techn., The Netherlands, Sept. 2004.
[127] P. van Zeijl, J.-W. T. Eikenbroek, P. P. Vervoort, S. Setty, J. Tangenberg, G. Shipton,

E. Kooistra, I. C. Keekstra, D. Belot, K. Visser, E. Bosma, and S. C. Blaakmeer, “A
Bluetooth radio in 0.18µm CMOS,” IEEE J. Solid-State Circuits, vol. 37, no. 12,
pp. 1679–1687, Dec. 2002.
[128] E. Hioe, K. Maio, T. Oshima, Y. Shibahara, T. Doi, K. Ozaki, and S. Arayashiki,

“0.18µm CMOS Bluetooth analog receriver with −88dbm sensitivity,” IEEE J.
Solid-State Circuits, vol. 39, no. 2, pp. 374–377, Feb. 2004.
[129] O. Bajdechi and J. Huijsing, “A 1.8-V modulator interface for an electret

microphone with on-chip reference,” IEEE J. Solid-State Circuits, vol. 37, no. 3,
pp. 279–285, Mar. 2002.
[130] J. Sauerbrey and R. Thewes, “Ultra low voltage switched opamp modulator
for portable applications,” in CICC Digest of Tech. Papers, May 2001, pp. 35–38.
[131] Y. Yang, A. Chokhawala, M. Alexander, J. Melanson, and D. Heser, “A 114-dB

68-mW chopper-stabilized stereo multibit audio ADC in 5.62mm 2 ,” IEEE J. Solid-
State Circuits, vol. 38, no. 12, pp. 2061–2068, Dec. 2003.
[132] M. Keskin, U.-K. Moon, and G. Temes, “A 1-V 10-MHz clock-rate 13-bit CMOS
modulator using unity-gain-reset opamps,” IEEE J. Solid-State Circuits,
vol. 37, no. 7, pp. 817–824, July 2002.
[133] E. Fogleman, J. Welz, and I. Galton, “An audio ADC -modulator with 100-dB
peak SINAD and 102-dB DR using a second-order mismatch-shaping DAC,” IEEE
J. Solid-State Circuits, vol. 36, no. 3, pp. 339–348, Mar. 2001.
[134] K. Nguyen and K. Sweetland, “A 105dB SNR multibit ADC for digital audio
applications,” in CICC Digest of Tech. Papers, May 2001, pp. 27–30.
[135] M. Dessouky and A. Kaiser, “Very low-voltage digital-audio modulator with

88-dB dynamic range using local switch bootstrapping,” IEEE J. Solid-State Cir-
cuits, vol. 36, no. 3, pp. 349 – 355, Mar. 2001.
REFERENCES 225
[136] A. Dezzani and E. Andre, “A 1.2-V dual-mode WCDMA/GPRS modulator,”

in ISSCC Digest of Tech. Papers, vol. 1, Feb. 2003, pp. 58 – 59.
[137] H. Aboushady, F. Montaudon, F. Paillardet, and M. Louerat, “A 5mW, 100kHz

bandwidth, current-mode continuous-time modulator with 84dB dynamic
range,” in ESSCIRC Digest of Tech. Papers, Sept. 2002, pp. 283–286.
[138] A. Nagari, A. Mecchia, E. Viani, S. Pernici, P. Confalonieri, and G. Nicollini, “A

2.7-V 11.8-mW baseband ADC with 72-dB dynamic range for GSM applications,”
IEEE J. Solid-State Circuits, vol. 35, no. 6, pp. 798 – 806, June 2000.
[139] O. Oliaei, P. Clement, and P. Gorisse, “A 5-mW modulator with 84-dB dy-
namic range for GSM/EDGE,” IEEE J. Solid-State Circuits, vol. 37, no. 1, pp.
2–10, Jan. 2002.
[140] V. Cheung, H. Luong, and W.-H. Ki, “A 1-V 10.7-MHz switched-opamp bandpass
modulator using double-sampling finite-gain-compensation technique,” IEEE
J. Solid-State Circuits, vol. 37, no. 10, pp. 1215–1225, Oct. 2002.
[141] P. Cusinato, D. Tonietto, F. Stefani, and A. Baschirotto, “A 3.3-V CMOS 10.7-

MHz sixth-order bandpass modulator with 74-dB dynamic range,” IEEE J.
Solid-State Circuits, vol. 36, no. 4, pp. 629 – 638, Apr. 2001.
[142] M. Miller and G. Petrie, “A multi-bit ADC for multi-mode receivers,” IEEE J.
Solid-State Circuits, vol. 38, no. 3, pp. 475–482, Mar. 2003.
[143] T. Burger and Q. Huang, “A 13.5mW, 185MSample/s -modulator for

UMTS/GSM dual-standard IF reception,” IEEE J. Solid-State Circuits, vol. 36,
no. 12, pp. 1868–1878, Dec. 2001.
[144] P. Rombouts, , J. D. Maeyer, and L. Weyten, “A 250kHz 94dB double-sampling

modulation A/D converter with a modified noise transfer function,” IEEE J.
Solid-State Circuits, vol. 38, no. 10, pp. 1657–1662, Oct. 2003.
[145] P. Maulik, M. Chadha, W. Lee, and P. Crawley, “A 16-bit 250-kHz delta-sigma

modulator and decimation filter,” IEEE J. Solid-State Circuits, vol. 35, no. 4, pp.
458 – 467, Apr. 2000.
[146] F. Esfahani, P. Basedau, R. Ryter, and R. Becker, “A fourth order continuous-time

complex sigma-delta ADC for low-IF GSM and EDGE receivers,” in VLSI Circuits
Digest of Tech. Papers, June 2003, pp. 75 – 78.
[147] T. Salo, S. Lindfors, T. Hollman, J. Jarvinen, and K. Halonen, “80MHz bandpass

modulators for multimode digital IF receivers,” IEEE J. Solid-State Circuits,
vol. 38, no. 3, pp. 464–474, Mar. 2003.
[148] T. Salo, S. Lindfors, and K. Halonen, “A 80MHz bandpass modulator for a

100-MHz IF receiver,” IEEE J. Solid-State Circuits, vol. 37, no. 7, pp. 798–808,
July 2002.
226 REFERENCES
[149] R. Schreier, J. L. ana L. Singer, D. Paterson, M. Timko, M. Hensley, G. Patterson,

K. Behel, and J. Zhou, “A 10-300-MHz IF-digitizing IC with 90-105-dB dynamic
range and 15-333-kHz bandwidth,” IEEE J. Solid-State Circuits, vol. 37, no. 12,
pp. 1636–1644, Dec. 2002.
[150] R. Gaggl, A. Wiesbauer, G. Fritz, C. Schranz, and P. Pessl, “A 85-dB dynamic

range multibit delta-sigma ADC CMOS bandpass continuous-time multibit -
ADC for ADSL-CO applications in 0.18µm CMOS,” IEEE J. Solid-State Circuits,
vol. 38, no. 7, pp. 1105–1114, July 2003.
[151] K. Philips, P. Nuijten, R. Roovers, A. van Roermund, F. M. noz, M. T. Pallares, and

A. Torralba, “A continuous-time ADC with increased immunity to interferers,”
IEEE J. Solid-State Circuits, vol. 39, no. 12, pp. 2170– 2178, Dec. 2004.
[152] S. Gupta and V. Fong, “A 64-MHz clock-rate ADC with 88-dB SNDR and
-105-dB IM3 distortion at a 1.5-MHz signal frequency,” IEEE J. Solid-State Cir-
cuits, vol. 37, no. 12, pp. 1653–1661, Dec. 2002.
[153] T.-H. Kuo, K.-D. Chen, and H.-R. Yeng, “A wideband CMOS sigma-delta mod-
ulator with incremental data weighted averaging,” IEEE J. Solid-State Circuits,
vol. 37, no. 1, pp. 11–17, Jan. 2002.
[154] I. Fujimori, L. Longo, A. Hairapetian, K. Seiyama, S. Kosic, C. Jun, and C. Shu-

Lap, “A 90-dB SNR 2.5-MHz output-rate ADC using cascaded multibit delta-
sigma modulation at 8 oversampling ratio,” IEEE J. Solid-State Circuits, vol. 35,
no. 12, pp. 1820 – 1828, Dec. 2000.
[155] T. Ueno and T. Itakura, “A 0.9V 1.5mW continuous-time modulator for

WCDMA,” in ISSCC Digest of Tech. Papers, Feb. 2004, pp. 78–79, 514.
[156] L. Dorrer, F. Kuttner, A. Wiesbauer, A. D. Giandomenico, and T. Hartig, “10-

bit, 3mW continuous-time ADC for UMTS in a 0.12µm CMOS process,” in
ESSCIRC Digest of Tech. Papers, Sept. 2003, pp. 245–248.
[157] R. Jiang and T. Fiez, “A 1.8 V 14 b - A/D converter with 4MSamples/s con-
version,” in ISSCC Digest of Tech. Papers, Feb. 2002, pp. 220 – 461.
[158] K. Vleugels, S. Rabii, and B. Wooley, “A 2.5-V sigma-delta modulator for broad-
band communications applications,” IEEE J. Solid-State Circuits, vol. 36, no. 12,
pp. 1887–1899, Dec. 2001.
[159] A. Tabatabaei and B. A. Wooley, “A two-path bandpass SigmaDelta modulator

with extended noise shaping,” IEEE J. Solid-State Circuits, vol. 35, pp. 1799 –
1809, Apr. 2002.
[160] F. Ying and F. Maloberti, “A mirror image free two-path bandpass modulator
with 72dB SNR and 86dB SFDR,” in ISSCC Digest of Tech. Papers, Feb. 2004, pp.
84–85, 514.
REFERENCES 227
[161] L. Luh, J. C. Jr., and J. Dapper, “A 400MHz 5th-order CMOS continuous-time

switched current modulator,” in ESSCIRC Digest of Tech. Papers, sep 2000,
pp. 72–75.
[162] Y. Geerts, M. Steyaert, and W. Sansen, “A 12-bit 12.5 MS/s multi-bit CMOS
ADC,” in CICC Digest of Tech. Papers, May 200, pp. 21 – 24.
[163] P. Balmelli and Q. Huang, “A 25MS/s 14b 200mW modulatorin 0.18µm

CMOS,” in ISSCC Digest of Tech. Papers, Feb. 2004, pp. 74–75, 514.
[164] P. Wambacq and W. Sansen, Distortion analysis of analog integrated circuits. Nor-
well, MA: Kluwer Academic Publishers, 1998, pp. 375–384.
[165] H. Samueli, “A low-complexity multiplierless half-band recursive digital filter de-

sign,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, no. 3, pp. 442–444,
Mar. 1989.

(The International Series in Engineering and Computer Science 874) Kathleen Philips, Arthur H. M. van Roermund (auth.) - ΣΔ A - D CONVERSION FOR SIGNAL CONDITIONING-Springer Netherlands (2006)

Uploaded by

Copyright:

Available Formats

(The International Series in Engineering and Computer Science 874) Kathleen Philips, Arthur H. M. van Roermund (auth.) - ΣΔ A - D CONVERSION FOR SIGNAL CONDITIONING-Springer Netherlands (2006)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(The International Series in Engineering and Computer Science 874) Kathleen Philips, Arthur H. M. van Roermund (auth.) - ΣΔ A - D CONVERSION FOR SIGNAL CONDITIONING-Springer Netherlands (2006)

Uploaded by

Copyright:

Available Formats

Σ∆ A/D CONVERSION FOR SIGNAL CONDITIONING

THE INTERNATIONAL SERIES IN ENGINEERING AND COMPUTER

ANALOG CIRCUITS AND SIGNAL PROCESSING

WIDE-BAND WIDTH HIGH-DYNAMIC RANGE D/A CONVERTERS

Arthur H. M. van Roermund

ISBN-10 1-4020-4679-0 (HB)

Printed on acid-free paper

All Rights Reserved

Printed in the Netherlands.

List of symbols and abbreviations ix

2 The signal conditioning channel 9

4 Power consumption in channel building blocks 37

5 Full-analog and full-digital conditioning channels 57

6.4.3 Power/performance analysis . . . . . . . . . . . . . . . . . . . . 98

7 Digitization of the inter-die interface 105

9 Conditioning ADCs for Bluetooth 139

9.4 FFB-ADC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

10 General conclusions 191

A Overview of published ADCs 193

B Power/performance relation of analog circuits 199

C Power/performance relation of digital ﬁlters 203

D Third-order distortion in analog circuits and ADCs 207

E Power consumption in a data interface 211

Symbol Description Unit

BW bandwidth of the wanted signal Hz

Symbol Description Unit

NTF noise transfer function

Figure 1.1: Electronic system including “signal-processing” units and some

Digital signal processing

Digitization of signal conditioning

Digitization of inter-die interfaces

Figure 1.3: Nomenclature with respect to the signal-conditioning channel

Baseband A/D conditioning channels

Continuous-time single-bit sigma-delta conversion

Power consumption as cost parameter

power-efﬁcient digitization of applications requiring less SNR. Especially the ﬁltering-

The signal conditioning channel

2.1 Generic communication channel

A generic communication channel is represented in ﬁg. 2.1. A message encoded in a

channel capacity= BW · log2 (1 + SNR) (2.1)

Figure 2.1: A generic communication channel as described by Shannon

2.2 Performance parameters

2.3 Conventional conditioning channels

Figure 2.2: Interference and noise sources in a communication channel

2.4.1 Technology advances

Figure 2.4: System demands in terms of signal bandwidth and resolution

2.4.2 System demands

• more information needs to be exchanged;

• higher quality is demanded by the customer;

and many more.

2.4.3 Advances in digital signal processing and analog circuit design

2.4.4 Digitization of the architecture

new requirements, to encompass more features, to provide multi-mode and multi-standard

UMTS Example [15], [16], [17], [18], [19]:

noise and for distortion.

For the transmitter, a zero IF architecture is suitable. Essentially, a pair of 3-bit

Figure 2.7: A general input spectrum and deﬁnitions

demand a higher performance conditioning channel and technology evolves in an unfa-

Figure 2.8: Deﬁnition of IM3