Low Power Design Techniques For Digital Logic Circuits

Low Power Design Techniques for Digital Logic Circuits
By
Yinshui Xia
BSc, MSc
Copyright by Yinshui Xia 2003
A thesis presented in partial fulfilment

of the requirements for the degree of
Doctor of Philosophy
Napier University
School of Engineering
March 2003
Declaration
I declare that no portion of the work referred in this thesis has been submitted in support
of an application of another degree, qualification or other academic awards of this or any other
university or institution of learning.
Yinshui Xia
Acknow ledgements
The research presented in this thesis was carried out under the supervision of Professor A.
E. A. Almaini, School of Engineering, Napier University. I sincerely thank him for offering me
an opportunity to study at Napier University, guiding my research to the logic synthesis world,
and encouraging me during this research. I am very grateful to Professor Almaini for his regular
comments and invaluable suggestions. I do appreciate his patience to teach the scientific writing
in English and correct my research papers and this thesis.
Thanks are due to Professor J. Hajto, my second supervisor, for his essential encouragement
of this research.
I would like to acknowledge the financial support of the School of Engineering, Napier University, Edinburgh. Without this support, this research work would not have been possible.
I would like to thank Dr Lingli Wang of Altera, a former member of the digital techniques
group, for his kindly support to install Linux operating system and great help to learn Linux
operating system and GNU C language programming environment when I started this research.
I also would like to thank Dr Alex Bystrov, also a former member of the group, at Newcastle
University for the valuable discussion of my research.
Thanks are due to other members of the digital techniques group, especially Mr Belgasem
Ali, Mr Khalid Faraj and Mr lVleng Yang, for their enjoyable working environment and various
discussions. Thanks are also due to my former supervisors and research partners in the different
period in my life. They are Prof. X. Wu of Ningbo University, Prof. T. G. Clarkson of King's
College London, Dr Nan Zhuang of Synopsys.
The time I have spent in the School of Engineering of the Napier University has been the
most challenging of my life. I would like to thank all staffs and PhD students in the School for
their kind helps. Finally, the special thanks are due to my wife Qiufen and my daughter Kankan
for their understanding and support during my research stUdy.
II
Abstract
With the rapid increase in the density and the size of chips and systems, area and power dissipation become critical concern in Very Large Scale Integrated (VLSI) circuit design. Low power
design techniques are essential for today's VLSI industry. The history of symbolic logic and some
typical techniques for finite state machine (FSM) logic synthesis are reviewed.
The state assignment is used to optimize area and power dissipation for FSMs. Two cost
functions, targeting area and power, are presented. The Genetic Algorithm (GA) is used to search
for a good state assignment to minimize the cost functions. The algorithm has been implemented
in C. The program can produce better results than NOVA, which is integrated into SIS by DC
Berkeley, and other publications both in area and power tested by MCNC benchmarks.
Flip-flops are the core components of FSMs. The reduction of power dissipation from flip-flops
can save power for digital systems significantly. Three new kinds of flip-flops, called differential
CMOS single edge-triggered flip-flop with clock gating, double edge-triggered and multiple valued
flip-flops employing multiple valued clocks, are proposed. All circuits are simulated using PSpice.
Most researchers have focused on developing low-power techniques in AND/OR or NAND
& NOR based circuits. The low power techniques for AND /XOR based circuits are still in
their early stage of development.
To implement a complex function involving many inputs,
a form of decomposition into smaller subfunctions is required such that the subfunctions fit
into the primitive elements to be used in the implementation. Best polarity based XOR gate
decomposition technique has been developed, which targets low power using Huffman algorithm.
Compared to the published results, the proposed method shows considerable improvement in
power dissipation. Further, Boolean functions can be expressed by Fixed Polarity Reed-Muller
(FPRM) forms. Based on polarity transformation, an algorithm is developed and implemented
in C language which can find the best polarity for power and area optimization. Benchmark
examples of up to 21 inputs run on a personal computer are given.
xv
Contents
Declaration
Acknowledgements
ii
List of Abbreviations
iv
List of Figures
vi
List of Tables
vii
viii
Abstract
Introduction
1.1
Motivations of low power design. . . . . . .
1.2
VLSI chip design approaches for low power
1.2.1
Sources of power dissipation . . . . .
1.2.2
Design approaches of low power VLSI systems.
1.3
Low power design flow . . . . . . . .
1.4
Low power digital design techniques
1.5
11
1.4.1
State assignment for low power Finite State Machines (FSMs)
11
1.4.2
Low power flip-flop design . . . . . . . . . . . .
12
1.4.3
Low power logic synthesis for FPRM functions
13
Outline
......................... .
iii
14
Conventional low power design techniques
16
2.1
Low power design approaches of FSMs . . . . . . . . .
16
2.1.1
Finite state machines and their representations
16
2.1.2
FSM power estimation
18
2.1.3
FSM power optimization .
24
2.2 Low power design of flip-flops ..
27
2.3
Power optimization of XOR gate based circuits
28
2.4
Genetic algorithm.
29
2.4.1
Overview
30
2.4.2
Representation
30
2.4.3
Initial population
32
2.4.4
Evaluation. . . .
32
2.4.5
Parent selection .
33
2.4.6
Crossover
33
2.4.7
Mutation
33
2.4.8
Elitism ..
33
State assignment for area and power optimization
35
3.1
Finite state machines (FSMs) . . . . . . . . . . .
35
3.2
Terminology and parameter calculation of FSMs
37
3.3
Cost functions
...............
3.4 State assignment using Genetic algorithm
3.5
44
46
3.4.1
Solution representation .
46
3.4.2
Evaluation.
48
3.4.3
Crossover
49
3.4.4
Mutation
50
3.4.5
Elitism ..
51
3.4.6
Outline of the algorithm
51
Experimental results . . . . . .
52
IV
3.6
4
Summary . . . . . . . . . . . . . . . . . . . . . . .
Differential CMOS single edge-triggered flip-flop
57
4.1
Differential CMOS single edge-triggered flip-flop.
58
4.2
Circuit parameter optimization
59
....... .
4.2.1
Delay
4.2.2
Circuit optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
60
61
4.3
Differential CMOS Single Edge-Triggered Flip-Flop with Clock-Gating (CG-SETFF) 62
4.4
Circuit simulations and power dissipation measurement.
65
4.5
PDP measurement . . . . .
69
4.6
Low power binary counters
70
4.6.1
4-bit binary counter
70
4.6.2
Binary twisted ring counters
72
4.7
Summary . . . . . . . . . . . . . . .
75
5 Differential CMOS double edge-triggered flip-flops

5.1
Previous work . . . . . . . . . . . . . . . . . . . . . .
5.2
Clock Chain Based Double Edge-Triggered Flip-Flop (CCB-DETFF) using single

latch
5.3
5.4
6
54
76
77
78
5.2.1
CCB-DET flip-flop
79
5.2.2
Circuit simulations
80
Pass Transistor Based Differential CMOS Double Edge-Triggered Flip-Flop (PTBDETFF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
83
5.3.1
Voltage scaling property for CCB flip-flops.
83
5.3.2
PTB-DETFF circuit
85
5.3.3
Circuit simulations
86
Summary . . . . . . . . .
89
Multi-valued flip-flop approach
90
6.1
Traditional Binary CP (BCP) controlled multiple valued master slave flip-flops in

series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
92
6.2
Multiple valued CP (MVCP) controlled multiple valued flip-flops in parallel
93
6.3
Circuit implementation
95
6.4
Experimental results
99
6.5
Summary .. . . . .
102
104
7 Low power XOR gate decomposition
7.1
Introduction
104
7.2
Definitions and terminology
105
7.3
Previous work . . . . . . . .
109
7.4
Proposed low power XOR gate decomposition
111
7.4.1
Theoretical results .
112
7.4.2
Proposed algorithm.
114
7.5
117
7.6
Summary . . . . . .
120
Power and area optimization of FPRM functions
121
8.1
Definition and terminology.
122
8.2
Related work . . . . . . . .
124
8.3
Polarity based area and power minimization of FPRM functions .
126
8.3.1
Power estimation . . . . . .
126
8.3.2
Area and power estimation
127
8.3.3
Polarity conversion
129
8.4
Proposed algorithms
131
8.5
134
8.6
Summary . . . . . .
139
Conclusions and Future work
140
9.1
140
Low power state assignment .
VI
9.2
Low power flip-flop designs
141
9.3
Power optimization of FPRM functions.
142
9.4
Future work . . . . . . . . . . . . . . . .
142
Publications
144
References and Bibliography
145
Disk Containing the Software
158
vii
List of Abbreviations
ADD Algebraic Decision Diagram
AND AND operation
BAL-CS Basic And Logic Combining Section
BCP Binary Clock Pulse
BDD Binary Decision Diagram
CAD Computer-Aided Design
CCB Clock Chain Based
CG Clock Gating
CK Clock Signal
CMOS Complementary Metal-Oxide-Semiconductor
C 2 MOS Clocked CMOS
CP Clock Pulse
DC Direct Current
DET Double Edge-Triggered
DETFF Double Edge-Triggered Flip-Flop
EPT Even Polarity Transformation
FF Flip-Flop
FSM Finite State Machine
FPGA Filed Programmable Gate Array
FPRM Fixed Polarity Reed-Muller
GA Genetic Algorithm
HDL Hardware Description Language
IC Integrated Circuit
lTV Input Threshold Voltage
LCS Logic Combining Section
LID Location InDicator
LSB Less Significant Bit
Vlll
MCNC Microelectronics Center North Carolina

MSB Most Significant Bit
MV Multiple Valued
MVCP Multiple Valued Clock Pulse
MVFF Multiple Valued Flip-Flop
NLS Negative-Level-Sampling
NMOS N-type Metal-Oxide-Semiconductor
NP Narrow Pulse
OBDD Ordered Boolean Decision Diagram
OPT Odd Polarity Transformation
PDA Personal Digital Assistant
PDP Power Delay Product
PLA Programmable Logic Array
PLS Positive-Level-Sampling
PMOS P-type Metal-Oxide-Semiconductor
PPRM Positive Polarity Reed-Muller
PTB Pass Transistor Based
RTL Register Transfer Level
SA Switching Activity
SC Sampling Circuit
SET-FF Single Edge-Triggered Flip-Flop
SOP Sum Of Products
SPICE Simulation Program with Integrated Circuit Emphasis
STG State Transition Graph
STT State 'Il'ansition Table
TSPC 'Il'ue Single Phase Clocking
VHDL Very large scale Hardware Description Language
VLSI Very Large Scale Integration
XOR EXclusive OR operation
ix
List of Figures
1.1
Moore Law
........................ .
1.2
Market of portable applications is growing very rapidly .
1.3
Battery capacity during the past 30 years
1.4 Charging and discharging for an inverter
1.5
Low power design flow
10
2.1
FSM structure model .
17
2.2
Monte-Carlo-based technique flow chart
19
2.3
Symbolic computation of the conditional transition probability matrix . . . . ..
22
2.4
FSM structure model (a) single-clock, flip-flop based FSM model; (b) clock-gating
version. . . . . . . . . . . . . . . . . . . . . . . . . .
25
2.5
Reducing the switching activity by inserting register
26
2.6
GA cycle. . . . . .
31
2.7
Crossover operator
33
3.1
State transition graph and state table of DK27 (a) STG; (b) STT
36
3.2
Conditional state transition probability of DK27
39
3.3
State probability labelled beside its state node and state transition probability
..........
labelled on each edge of DK27 . . . . . . . . . . . . . . . . . . . . . . . . . . . ..

3.4
4.1
41
Relationship between cost function and optimization quality (a) Area and SA
versus a in equation ??; (b) Area and SA versus (3 in equation ??
53
Differential CMOS SET-FF (a) Circuit implementation; (b) Logic symbol
59
4.2
Power delay product optimization .
61
4.3
Power versus width w
62
4.4
Delay versus width w .
63
4.5
PDP versus width w .
64
4.6
Non-redundant transition clock chain.
64
4.7
Differential single edge-triggered CMOS flip-flop with clock-gating (CG-SETFF)(a)

Circuit implementation; (b) Logic symbol
65
4.8
Transition behavior of proposed flip-flop .
66
4.9
Test bench for measuring the power dissipation of flip-flops
67
4.10 Power dissipation against D input switching activity. PI: Proposed flip-flop; P2:
Conventional flip-flop . . . . . . . . . . . . . . . . . .
68
4.11 Timing test circuit (a) Bench circuit; (b) Timing diagram
69
4.12 PDP versus switching activities
70
4.13 SET-FF based binary counter.
72
4.14 Gate circuit implementation (a) 2 input NAND and its logic symbol; (b) 3 input
NAND and its logic symbol; (c) XNOR and its logic symbol. . . . . . . . . . ..
73
4.15 Power dissipation for twisted ring counters: P1- SET-FF based counters; P2-CGSETFF based counters . . . . . . . . . .
74
4.16 State diagram for a twisted ring counter
74
5.1
Differential CMOS DET flip-flop proposed in [68]: BALCS-DETFF
77
5.2
Differential CMOS single-edge triggered flip-flop (SET- FF) . . . .
78
5.3
Double edge pulse generation (a) Clock chain; (b) Timing diagram
80
5.4
Clock Chain Based double Edge-Triggered Flip-Flop (CCB-DETFF) cell (a) Flipflop cell; (b) Logic symbol . . . . . . . . . . . . . . . . . . .
5.5
Transient behavior (a) Differential CMOS SET-FF in Fig.??; (b) Differential

CCB-DETFF in Fig. ?? . . . . . . . . . . . . . . . . . . . .
5.6
80
82
Power dissipation versus switching activities: P1- result for Fig. ??; P2-result for
Fig. ??; P3-result for the proposed circuit . . . . . . . . . . . . . . . . . . . . ..
Xl
83
5.7
Input threshold voltage versus gate voltage of clocked transistors. NB: Vth-Two
stands for threshold voltage of Cir2ct while Vth-One for that of Circt .
84
5.8
Narrow pulse generating scheme. . . . . . . . . . . . . . . . . . . . . .
85
5.9
Proposed differential CMOS double edge-triggered flip-flop (DET-FF) (a) Narrow

pulse generator; (b) Double edge-triggered flip-flop cell; (c) Logic symbol.
86
5.10 Transition behavior of the propose circuit . . . . . . . . . . . . . . . . . .
87
6.1
Comparison between binary and quaternary signal transmission.
(a) Parallel
transmission; (b) Serial transmission . . . . . . . . . . . . . . . . . . . . . . . ..

6.2
91
BCP controlled quaternary master slave D flip-flop in series and two digit quaternary shift register (a) BCP controlled quaternary master slave D flip-flops in
series; (b) Two digit quaternary shift register; (c) Logic symbols . . . . . . . . ..
6.3
93
Logic structure of MVCP controlled multiple valued flip-flop in parallel (a) Logic
circuit; (b) Logic symbols . . . . . . . . . . . . . . . . . . . . . . . .
94
6.4
Threshold 1.5 inverter (a) Circuit implementation; (b) Logic symbol
98
6.5
Circuit implementation (a) Quaternary 2 to 1 multiplexer; (b) Quaternary reshaper; (c) Threshold function generator; (d) Quaternary 4 to 1 multiplexer . . . 100
6.6
PSpice simulation of the proposed flip-flop .
101
7.1
Switching activity versus signal probability
7.2
(a) Multi input XOR gate; (b) Published decomposition in [131] and [74]; (c) A
. . . . . . . . . . . . . . . . . . . . . 108
new decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

8.1
(a) Multi input AND gate decomposition in [107]; (b) Multi input XOR gate
decomposition in [131] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
8.2
Power dissipation with the function forms (a) The minimal form under polarity
4; (b) The function form under polarity 5 . . . . . . . . . . . . . . . . . . .
129
8.3
Decomposition and AND- trees and XOR tree (a) AND trees; (b) XOR tree
133
8.4
Power and area versus alpha (a) Power versus alpha; (b) Area versus alpha
138
xii
List of Tables
1.1
Some commercial and academic low power design tools
3.1
Tav's kiss format . . . . . . . . . . . . . . . . . . . . . . .
40
3.2
Experimental results showing power & area improvement
55
3.3
Best state assignment codes by GA: Assignments a & b are based on costa & costb
respectively . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
4.1
SPICE parameters of MOS transistors for a generic 1.0p,m process
65
4.2
Power dissipation for load capacitors versus switching activities
68
4.3
State table for a 4-bit binary counter . . . . . . . . . . . . . . .
71
5.1
Power savings compared to the circuit in Figs.?? &?? . . . . .
81
5.2
Power dissipation comparison between SET-FF and PTB-DETFF .
87
5.3
Voltage scaling property between SET-FF and PTB-DETFF . . . .
88
5.4
Voltage scaling property between BALCS-DETFF and PTB-DETFF
88
5.5
Power dissipation from clock chain and flip-flop cell with VD D = 5.0V and
Vclock
5.0V . . . . . . . . . . . . . . . . . . . . . . . .
89
6.1
SPICE parameters for a generic 0.5p,m process
99
6.2
Performance measurement of the MVFF (ns)
6.3
Area and Power dissipation for the proposed circuit and the circuit in Fig. ??(b)
102
7.1
Best polarity and switching activity for low power decomposition . . . . .
118
7.2
Experimental results for the proposed algorithm and published algorithms
119
xiii
101
8.1
Best polarity and switching activity for low power decomposition[127j
136
8.2
Reduction ratios of area and switching activity
136
8.3
Power dissipation and area comparison
137
8.4
Reduction ratios of area and SA . . . .
137
XlV
Chapter 1
Introduction
1.1
Motivations of low power design
The genesis of low power microelectronics can be traced to the invention of the transistor in 1947.
It was a breakthrough of virtually unparalleled importance in electronics to eliminate the crushing
needs for several watts of heater power and several hundred volts of anode voltage in vacuum
tubes in exchange for transistor operation in the tens of milliwatts range. The capability to fully
utilize the low power assets of the transistor was provided by the invention of the integrated
circuit in 1958. Since the first integrated circuit (IC) was developed by Jack Kilby in Texas
Instruments and then became the first commercial IC by Fairchild Instruments in 1961, the IC
technology has progressed greatly. The size and density of IC chips and systems are increasing
as Gordon Moore, co-founder of Intel, predicted in 1960s that the number of transistors on an IC
chip could be doubled every 12 to 18 months. This is well confirmed by the development trace
of Intel microprocessors as shown in Fig. 1.1.
The continuing increase in chip density and operating frequency have made power consumption a major concern in very large scale integrated (VLSI) circuit design. For example, the PC
chip from Motorola consumes 8.5W, the Pentium chip from Intel consumes 16W, and DEC's
21164 (300MHz on a die area of 3cm 2 ) consumes 50W! It is extrapolated that 10cm 2 microprocessor, clocked at 500MHz would consume 315 W in near future [83]. Unless power consumption
is dramatically reduced the resulting heat will limit the feasible packing and performance of VLSI
circuits and systems.
1
CHAPTER 1.
INTRODUCTION
Perhaps the primary driving factor for designing low power systems has been the remarkable
success and growth of the class of portable personal computing devices and wireless communication systems which demand complex functionality and high speed communication. Probably no
segment of the electronics industry has a growth potential as explosive as that of the personal
digital assistant (PDA) which has been characterized as a combined pocket cellular phone, pager,
e-mail terminal, fax, computer, calendar, address directory, notebook, etc. [96][69][26].
The market of portable applications is growing very rapidly. Fig. 1.2 shows the various PC
percentages of the PC market in 1992 and 1998 [60]. In portable applications, average power
consumption is a critical design constraint since it is related to the battery life time, size and
weight [55][66]. The reason for this is illustrated with the simple example of a multi-media terminal. The projected power for such a terminal, when implemented using off-the-shelf components
not designed for low-power operation [95], is about 40W. With advanced Nickel-Metal-Hydride
battery technologies yielding around 65 watt-hours/kilogram [56], this terminal would require an
unacceptable 6 kilograms of batteries for 10 hours of operation between recharges. Even with new
Number of Transistors
228 1
Intel Itanium 2(1GHz) ~

Pentium III Xeon(866 MHz)~
Mobile Pentium II(400 MHz) ~
Intel Ce1eron (333 MHz)~
226 I224 I-
Pentium Pro (200 MHz) ~

Pentium (66 MHz) ~
222
220
Intel486 DX ~
I-
218 I-
Inte1386 DX~
80286 ~
216 I8086~
214 ,
212 l- 800~080~
TO
140~~
~I
______~________~______- L_ _ _ _ _ _~_ _ _ _ _ _ _ _L -_ _ _ _ _ _~_ _ _ _ _ _~
1970
1975
1980
1985
1990
Introduction Year
* These data are originally from
1995
2000
"I1ltel Microprocessor Quick Refere1lce Guide"
Figure 1.1: Moore Law
2005
CHAPTER 1.
INTRODUCTION
battery technologies, such as rechargeable Lithium Ion or Lithium Polymer cells, it is anticipated
that the expected battery lifetime will increase to about 90-110 watt-hours/kilogram, which still
leads to an unacceptable 3.6 to 4.4 kilograms of battery cells. Fig. 1.3 shows the battery capacity
increasing over the last 30 years [56]. From Fig. 1.3, it can be observed that battery capacity has
only improved with a factor 2 to 4 over the last 30 years while the computation power of digital
IC's has increased by more than 4 orders of magnitude [51]. The gap is increasing with respect to
power demand. In the absence of low power design techniques, the current and future portable
devices will suffer from either very short battery life or unreasonably heavy battery pack.
1992 PC Market
1998 P( Mark et
Mobile
Dertop 70%
Desktop 49%
Figure 1.2: Market of portable applications is growing very rapidly

These problems make it necessary to develop power aware VLSI design tools that help achieve
low power in these systems. Indeed, the Semiconductor Industry Association has identified lowpower design techniques as a necessary technological need [93].
1.2
1.2.1
VLSI chip design approaches for low power

Sources of power dissipation
Since CMOS technology is predominant in the realization of today's IC and CMOS devices are
intrinsically low power consuming, all circuits for the rest of this thesis refer to CMOS circuits.
It is judicious to briefly discuss the mechanisms for power consumption in CMOS circuits. Take
the inverter in Fig. 1.4 for example. The power dissipation for the inverter consists of three
INTRODUCTION
CHAPTER 1.
major sources which are expressed in the following equation:
Ptotal
The first term,
Pleakage,
= Pleakage + PshorLcurrent + Pswitching
(1.1)
can arise from substrate injection and subthreshold effects and is pri-
marily determined by fabrication technology considerations [24]. The second term,
PshorLcurrent,
is due to the direct-path short circuit current, which arises when both the NMOS and PMOS
transistors are simultaneously active, conducting current directly from supply to ground [118].
The third term,
Pswitching,
represents the switching power dissipation or dynamic power dissipa-
tion. This is the result of capacitance charging and discharging in the circuit. The situation is
modeled in Fig.1.4 where the parasitic capacitances are lumped at the output in the capacitor
CL. Consider the behavior of the circuit over one full cycle of operation with the input voltage
going from
to ground and back to
Vdd
:2
Vdd'
When the input changes from
to ground, the
Vdd
40
--..
i'/J
L..
:::::;
30
n e
+"
+"
o::i
Nickel-Cadm'lum
o::i
Q..
o::i
10
l
l
o::i
E
o
In
I
, I
>- 20
;!:::
!
,
o
65
70
75
80
85
90
Year
Figure 1.3: Battery capacity during the past 30 years
95
CHAPTER 1.
capacitor
CL
INTRODUCTION
is charged. This charging process draws an energy equal to CL
supply and ~CL
Vld from the power
Vld is stored in the capacitor CL because at the end of the transition the output
capacitor CL is charged to Vdd. Hence, ~CL Vld is dissipated in the PMOS network. When the
input changes from ground to Vdd, the capacitor CL is discharged. And ~CL Vld is dissipated in
the NMOS network. Hence, for a full cycle,
CL
Vld from the power supply is consumed.
This
leads to the conclusion that CMOS power consumption depends on the switching activity of
the signals involved. If a represents the signal activity, the expected number of zero to one
transitions per data cycle, and f is the average data-rate, which may be the clock frequency in
a synchronous system, then the effective frequency of nodal charging is given by the product of
the activity and the data rate:
af.
Hence, the average CMOS power consumption is given by
equation 1.2.
Pswitching
(1.2)
2aCL Vddf
Vdd
SL
Figure 1.4: Charging and discharging for an inverter
For a well designed circuit, the first two terms in equation 1.1 can be kept below 20% of
the total power [14]. Hence, in CMOS circuits, Pswitching is by far the most important. From
equation 1.2, Pswitching is proportional to the switching activity, capacitance loading, data-rate
(in synchronous systems fmight correspond to the clock frequency), and the square ofthe supply
voltage.
CHAPTER 1.
1.2.2
INTRODUCTION
Design approaches of low power VLSI systems
From the above analysis of power dissipation sources, the vast majority of power reduction
techniques concentrate on minimizing the dynamic power dissipation by reducing one or more
factors on the right hand side of equation 1.2 .
Voltage scaling
One of the most obvious ways to reduce power is to reduce the power supply voltage of circuits
because the dynamic power is proportional to the square of the supply voltage. Power savings
are relatively independent of circuit function and circuit technology. Hence, it is applicable at
different stages of the design development. Some methodologies have been proposed [24][25].
However, with the scaling of supply voltage and device dimensions, the transistor threshold
voltage also has to be scaled to achieve the required performance. Unfortunately, such scaling
does not come for free and can increase the leakage current. The leakage current occurs due
to carrier diffusion between the source and the drain when the gate-source voltage, Vgs , has
exceeded the weak inversion point, but is still below the threshold voltage 11th, where carrier
drift is dominant. The current in the subthreshold region is given by equation 1.3[24]
Isub
= kexp(
v;,s - 11th
n
1 1 1 f"\
Vds
)(1 - exp( - VT))
(1.3)
Where k is a function of the technology, VT is the thermal voltage (KT / q), 11th is the threshold
voltage and S is the subthreshold swing.
Due to the exponential nature of subthreshold leakage current with 11th, subthreshold current
can no longer be ignored. The lower the 11th is, the higher the subthreshold leakage current will
be. On the other hand, delay through a logic block is proportional to Vdd/ (Vdd - 1Ith)2 where
Vdd is the voltage of the power supply [24][25]. Lower 11th means longer delay. To compromise
the two, it is suggested to use high supply voltage in the critical paths of a design to achieve
the required performance while the off-critical paths of the design use lower supply voltage to
achieve low-power dissipation [53][28]. It is also clear from the above discussion that reducing
CHAPTER 1.
INTRODUCTION
the threshold voltage allows the supply voltage to be scaled down to lower Pswitching without loss
in speed. However, the limit on threshold voltage scaling is imposed by the noise margin and the
increase of sub-threshold current. Scaling down threshold voltage trades off between dynamic
power (Pswitching) and static power (PLeakage)' In considering logic and memory circuit behavior,
Vdd = 1.0V appears to be a good compromise for small dynamic and static power dissipation
[89]. A genius approach is architecture-driven supply voltage scaling based strategy which is to
modify the architecture of the system so as to make it faster and reduce Vdd so as to restore
the original speed resulting in reduced power consumption [24][25]. However, area overhead is
required.
Frequency reduction
The second obvious way to reduce power is to decrease the clock frequency
f. Decreasing f
causes a proportional decrease in power dissipation. However, slowing the clock will result in
a slower computation. The power consumption over a given period of time is reduced, but the
total amount of useful work is reduced as well. As a result, the energy dissipated to complete the
task has not changed. This can be illustrated by an example. Suppose the system is clocked with
a clock period TI , and the task takes N clock cycles to complete. During each cycle, the system
dissipates an average power Pl. If the frequency is decreased by half, the power dissipation
over the original time period will be P 2 = ~ PI because average power is directly proportional
to the clock frequency. However, it now takes a total time of 2NTI to complete the task. As a
consequence, the average energy consumed by the system is E
= PINTI in both cases.
However,
this observation still incites some genius low power designs. One of them is double edge-triggered
flip-flops targeting low power [49][19][68][102], which enables a halving of the clock frequency and
hence reduces power dissipation on the clock line for a given data rate compared with the single
edge-triggered flip-flops .
Effective capacitance reduction
Effective capacitance, CEff' is defined as C Eff
aCL. Design and synthesis techniques have
been developed to reduce both the capacitive load, CL, and the switching activity, a, at all stages
CHAPTER 1.
INTRODUCTION
ofthe design process[37][45][65] [27][61] [78] [88][108][131]. It is obvious that once a technology and
a supply voltage have been set, power savings come from the careful minimization of the effective
capacitance. There are many effective applications of this idea. Some examples are shown as
follows.
1. Dynamic power management: The sleep modes of operation in portable computers [37][45]
are examples of this approach to reduce useless switching activity. Power is reduced by
stopping the clock or shutting down the power supply of parts of the system that are not
required to carry out the current task.
2. Algorithmic transformations for signal processing tasks [65][27]: Reducing the number of
operations needed to carry out a given computation may not be always useful in terms of
performance, but it is often useful for reducing power.
3. Communication protocol design [61]: Communication protocols can be modified to improve
the activity patterns.
4. Memory allocation [78] and bus encoding [98] techniques to minimize the power dissipated
in memories and system busses.
5. Logic optimization [88][108][114]: Area minimization can obtain global power savings because area is directly related to the capacitive load. Logic optimization can directly target
power optimization.
6. Circuit topology [131][74]: This method aims to change circuit topology to minimize the
switching activity of a circuit so that the effective capacitance is minimized.
Other approaches
A revolutionary approach for low power design is called energy recovery jrecycle techniques, which
are addressed in [11][34]. Scaling down feature size is an important issue for high-performance
and high-density VLSI circuits. However, some second order effects become serious and are
becoming a major challenge in deep submicrometer devices and circuits. Corresponding low
power techniques are being explored [52].
CHAPTER 1.
1.3
INTRODUCTION
Low power design flow
Designing for low power is at least as difficult as designing for maximum speed or minimum
area. Power dissipation is a pattern-dependent cost function, unlike area, which is constant with
respect to input patterns. Since power dissipation becomes increasingly important as a design
evaluation metric, a new generation of computer-aided design tools targeting power minimization
is urgently needed by designers. In the last few years, significant research and development efforts
have been undertaken in academia and industry targeting the creation of a new generation of
computer-aided design (CAD) tools for low power. As a result, hundreds of papers have been
published on the subject [88][104][12][108][77]. The wide range of ECAD tools for low power fall
into four major categories based on the four levels, which are behavioral, architectural (Register
Transfer/RT), logic and circuit level [89]. Countless commercial and academic design synthesis
tools are available and some of them are shown in Table 1.1. It can be seen that VHDL [81] and
Verilog [103] are the most popular hardware description languages (HDLs).
Table 1.1: Some commercial and academic low power design

Input
Organization System
Description
Synopsys
PowerMill
Circuit
Simula- SPICE, Verilog
tion
Power
RTL Power Esti- Gate
Synopsys
Netlist,
Compiler
mati on
Switching Activity, Constraints
PowerHogs
RTL Power Opti- Verilog/VHDL
Synopsys
mization
PowerSim
Cadence
Gate and Logic Verilog
Level Estimation
TransEDA
PowerSure
HDL/RTL Power Verilog/VHDL
Estimation
Lsim Power Circuit level Sim- SPICE
Mentor
Analyst
ulation
Power-Tool Logic level Power Verilog
Veritools
Estimation
Berkeley
University
SIS
Logic Level, RTL

Power Estimation
Gate Netlist
tools
Environment
Unix
Unix
Unix
Unix
Unix
Workstation
Win95,
WinNT,
Sun4.x
Win95,
Unix
A typical synthesis based VLSI low power design flow is shown in Fig. 1.5[60].
CHAPTER 1.
INTRODUCTION
10
It can be seen from Fig. 1.5 that three steps are included in low power synthesis at each
design level:
1. Estimate power according to unoptimized description. The methods of power estimation
are different from level to level. This process is called power analysis.
2. Optimize the description through various available procedures by the criteria of power
dissipation, area, speed or testability. This important process is called optimization.
trnnsfommtion
Behavioral
Power Analysis
Logic Synthesis
and
Optimization
RTL
Mapping
BB
EJEJ
Tee-Mapping
and
Circuit-level
Optimization
Logic-level
Power Analysis
Circuit-level
Power Analysis
Figure 1.5: Low power design flow
CHAPTER 1.
INTRODUCTION
11
3. Produce power optimization description.

Step 1 and Step 2 are very important for low power design. Accurate power estimation tools
must provide feedback on the quality of each design choice. In the last ten years, at each level of
abstraction, various design alternatives have been explored. The availability of a power estimator
for each level of abstraction is fundamental in a low-power design flow. To avoid costly re-design
steps, it is mandatory to be able to optimize the power dissipation during the early stages of the
design process.
1.4
Low power digital design techniques
Low power digital design techniques cover a broad range of subjects. In this thesis, three topics
are chosen to discuss.
1.4.1
State assignment for low power Finite State Machines (FSMs)
Most of VLSI circuits are sequential circuits. A general model of sequential circuit structures is
called Finite State Machine (FSM), which is composed of two sections: Combinational section and
register section. Register section usually is composed of flip-flops. Compared to combinational
circuits, there are two working characteristics:
An FSM has flip-flops to store state signals.
An FSM has clock signals to synchronously trigger fiip-flops and to realize the synchronous
switching of state variables.
Thereby, its synthesis procedure is slightly different from combinational circuits. FSM synthesis
can be divided into four stages[15]:
Behavior synthesis: Obtain representation of machine behavior expressed as State Transition Tables (STTs) or State Transition Graphs (STGs).
State assignment: Assign unique binary code to the symbolic states of an FSM and obtain
a description of the circuit produced in terms of Boolean functions.
CHAPTER 1.
INTRODUCTION
12
Logic synthesis: Optimize Boolean functions with respect to the original cost metric.
Library binding: Map the optimization Boolean functions to components from a standard
gate library and produce a gate-level description.
State assignment is the critical step in low power design of FSMs and is one subject of this
thesis. It has an important role in determining the number of nodes required to implement the
output and next logic functions. The reduction of the number of nodes promises global power
savings. On the other hand, state assignment directly determines the switching activities of the
state variables and the interior variables in the combinational circuit.
The contribution of this work is in the formulation of the problem that links switching
activities of an FSM to its power dissipation and in the study of Genetic Algorithms (GAs) for
the search of optimal solutions to the problem of finding a state assignment that gives low power
dissipation. Only minimizing switching activities on the state lines in the FSM does not guarantee
the reduction of the total power dissipation because the power consumed in the combinational
section is not taken into account. Instead, more accurate cost functions which take that of the
combinational section into account have been developed. A methodology and a software package
that combines ESPRESSO[23] are developed. Test results on MCNC benchmark circuits show
that our package performs significantly better than other synthesis tools [94][111] in the majority
of cases.
1.4.2
Low power flip-flop design
Flip-flops are the core components of FSMs. Reducing power dissipation of flip-flops can result
in significant power reduction of FSMs. Traditional flip-flops are single-edge triggered flip-flops
(SETFF), which are sensitive to the rising or falling edge of the clock. Narrow pulse sampling
based single edge-triggered flip-flops are implemented using a clock chain to generate a series
of narrow pulses. It suffers from significant redundant transitions in a long clock chain, which
results in redundant dynamic power dissipation. To solve this problem, a clock-gating scheme is
proposed to eliminate the redundant transitions and a low power single edge-triggered flip-flop is
presented. On the other hand, for single edge-triggered flip-flops, half of the clock's transitions
INTRODUCTION
CHAPTER 1.
13
are redundant, which results in wasteful dynamic power dissipation. To improve this, doubleedge triggered flip-flops (DETFFs) are proposed, which utilize both transition edges of the clock,
achieving power savings. In this thesis, a low power SETFF is proposed, in which clock-gating
techniques are used to reduce the redundant transitions of clock signals for saving power and the
structure of low power DETFFs are explored and multiple valued flip-flops are investigated to
explore a novel solution of low power flip-flop design.
1.4.3
Low power logic synthesis for FPRM functions
Any n-variable Boolean function! can be expressed by Shannon expansion based on AND/OR
operation as follows.
!(X n -1X n -2'" xo)

where 0
= XdXi=O + XdXi=l
(1.4)
n - 1, and !x;=o and !x;=l are the cofactors of ! with respect to Xi. Corre-
spondingly, a broad range of logic minimizers are available for SOP forms such as ESPRESSO
[23] and SIS[94].
Alternatively, any Boolean function can be represented by AND /XOR operations, which is
called Reed-Muller expansion.
!(X n -1X n -2'" xo)
= XdXi=O E9 XdXi=l
(1.5)
!(X n -1X n -2' .. xo) = !Xi=O E9 Xi (fx;=o E9 !Xi=l)
(1.6)
!(X n -1Xn -2'" XO) = !Xi=l E9 Xi(fxi E9 !xi=d
(1.7)
CHAPTER 1.
INTRODUCTION
14
In logic synthesis, Reed-Muller logic methods are important alternatives to the traditional
SOP approaches to implement Boolean functions. Reed-Muller realizations have some attractive
advantages especially for functions that do not produce efficient solutions using SOP techniques.
In addition, XOR based circuits have great advantage of easy testability. However, due to
the lack of efficient conversion tools and Reed-Muller logic optimizer, applications of ReedMuller implementations have not become popular. With the development of FPGAs, XOR gates
are already manufactured as basic cell components, which encourages the research on ReedMuller logic optimization. There has been extensive research on Reed-Muller methods targeting
area minimization [4][6][8][79] [64][114]. However, the research on low power Reed-Muller logic
implementations is still in their early stage of development.
XOR Gate decomposition is the step before technology mapping, which decompose a multi
input XOR gate into a two input XOR gate tree. Low power XOR gate decomposition is one
solution of low power Reed-Muller logic implementations. Based on polarity conversion, a novel
XOR gate decomposition targeting low power is proposed.
Any Boolean function can be represented canonically by a Fixed Polarity Reed-Muller (FPRM)
form. Minimization of FPRM functions promises global power savings. A power estimation frame
for FPRM functions is proposed. Based on polarity conversion, a power minimization algorithm
is developed.
1.5
Outline
This thesis covers three main parts: Low power state assignment for FSMs; low power flipflop design and power optimization for FPRM functions. The conventional low power design
techniques are reviewed in chapter 2. The first part investigates low power state assignment for
FSMs. The problem is formulated in chapter 3. Two cost functions are proposed, which take the
power of the combinational section into account. Genetic Algorithm combined with ESPRESSO
is used to search the optimal solution. The second part is for low power design of flip-flops. A
new type of low power single-edge triggered flip-flop is proposed in chapter 4 while double-edge
triggered flip-flops are presented in chapter 5. A multiple valued approach to design low power
CHAPTER 1.
INTRODUCTION
15
flip-flops is explored in chapter 6. The third part deals with Reed-Muller logic which is based
on AND /XOR operations. Based on polarity searching, low power XOR gate decomposition
is stated in chapter 7 while low power FPRM function minimization is described in chapter
8. Finally, the main improvements and contributions are summarized and some future work is
suggested in the "conclusions and future work".
Chapter 2
Conventional low power design

techniques
2.1
Low power design approaches of FSMs
Most of VLSI circuits are sequential ones. A sequential function can be represented by several
models [50]. Usually, it is modeled by a finite state machine (FSM) [67].
2.1.1
Finite state machines and their representations
Low power design of FSMs involves tackling the problem of the power estimation method and
power optimization strategy. They can be done at each of four design levels [15].
An FSM is defined by the following standard definition.
Definition 2.1. An FSM is characterized by a 5-tuple (X, Y, S, A, ry) where X, Y, S are the
sets of primary inputs, primary outputs and internal states and A, ry are the output and next
state functions, respectively. The FSM is represented by a state transition table (STT) M
{mdml
(Xl, (Sdl, (Si)I,YI),l E {O,1, .... L -1}), here, L is the number of product terms, Xl is
the primary input, Si,
si
S are the present state and the next state and YI is the corresponding
output. Each entry ml E M is a symbolic implicant of the FSM.

FSMs are categorized in two classes [14]:
16
CHAPTER 2. CONVENTIONAL LOW POWER DESIGN TECHNIQUES

Definition 2.2. A Moore machine is an FSM where .\(x, s)
= .\(s),
17
i.e., the outputs do not
depend directly on the input value, but they depend only on the state. A Mealy machine is an
FSM for which this property does not hold.
FSMs can be incompletely specified. An incompletely specified FSM is one where 'f](x, s)
and / or .\(x, s) are incompletely specified Boolean functions. An FSM can be represented by
a graph or, equivalently, by a table. The two representations are called state transition graph
(STG) and state transition table (STT), respectively. The states of the STG are labeled with
the unique symbolic state name. The edges are labelled with the input and output values. The
state table is simply the list of edges of the STG.
Both STG and STT completely define the input-output behavior of an FSM, but they do not
provide any information about the circuit implementation. Hence, STG and STT are behavioral
representations of the FSM. In order to obtain a representation which is closer to the circuit
implementation, the concept of state encoding is needed to be introduced.
Definition 2.3. A state encoding is a one-to-one mapping from S to B N (Boolean space), i.e.,
a function E: S
--r
B N . The number of state variables is indicated by Ns and Ns ~
rlo921S11.
Once Ns and E have been specified, the state of an FSM is completely expressed by Ns binary
variables called state variables. Once the state encoding has been specified, the structural model
of an FSM is shown in Fig. 2.1. The representation of Fig. 2.1 is structural, because it refers to
a particular circuit structure implementing the FSM.
Primar~
Primary
outputs
inputs ~I Combinational
logic
Present-s tate
Next-state
inputs
CP_ _ _ _-'
Figure 2.1: FSM structure model
The structural representation is useful since it may be much more compact than the state-
18
based representations like STT and STG. However, the main limit of structural representations
is that they are not unique.
Binary decision diagrams (BDDs) are a data structure developed for the compact representation of large Boolean functions. Several variants of BDDs have been developed by different
groups of researchers [22][20]. It has been shown that the Reduced Ordered BDD (ROBDD) is a
canonical form, i.e., two functions are equivalent if and only if they have the same BDD. There
are some distinct advantages to represent an FSM by BDD [39].
2.1.2
FSM power estimation
The process of IC design involves a transformation of a high level behavioral specification to a

lower level architectural (or RTL) specification, and then to a lower gate-level specification, and
so on. If one were to get to the transistor or gate level design and only then discover that the
power consumption is unacceptably high, it would be too expensive to make design changes. The
circuit may require significant rework, involving perhaps changes to the overall architecture of
the chip. For this reason, it would be very beneficial to have a power estimation capability at
a high level of abstraction. However, estimation from a high level of abstraction is potentially
inaccurate, while low-level power estimation can be very accurate. Therefore, a power estimation
capability is needed at every level of abstraction in order to check the design at every step .
SPICE
The acronym SPICE stands for Simulation Program with Integrated Circuit Emphasis [86]. It
is a general-purpose circuit program that simulates electronic circuits and can perform various
analysis of electronic circuits: the operating points of transistors, a time-domain response, a
small-signal frequency response, and so on. SPICE contains models for common circuit elements,
active as well as passive, and it is capable of simulating most electronic circuits. It is a versatile
program and is widely used by industries and universities. The main frame versions are HSpice
(Meta-Software), PSpice (MicrSim), AccuSim (Mentor Graphics) and Cadence-SPICE (Cadence
Design). The PC-version, PSpice (MicroSim), is also available. PSpice allows various types of
analysis, which include DC Sweep, Transient Analysis and AC Analysis. For the average power
19
estimation of a circuit, the library function avg(i x v) can be called where i is the node current
and v is node voltage. However, the power dissipation is input pattern-dependent. Hence, the
utility of straightforward simulation can be limited because it is time consuming and can only
work at transistor level.
Monte-Carlo-based power estimation
The basic idea of Monte Carlo methods for estimating activities of individual nodes is to simulate
a circuit by applying random-pattern inputs. The convergence of simulation can be obtained
when the activities of individual nodes satisfy some stopping criteria. The procedure is outlined
in Fig. 2.2 . The detail is discussed in [73].
Generate a Random Circuit State
Generate Inputs (a,P) and Sampl
No
Figure 2.2: Monte-Carlo-based technique flow chart
Encoding based power estimation

This targets the STT or STG level of abstraction. At the STT or STG level of abstraction,
nothing has been decided about the structure of the combinational logic implementing the next
state and output functions. The state assignment algorithm can exploit degrees of freedom that
20
are lost at successive phases and produce an encoded state transition table that is an effective
starting point for further power optimizations of the combinational logic. Some power estimation
methods have been developed at this level.
Power is a strongly pattern-dependent cost function, since it depends on the switching activity
of a circuit, which in turn depends on the input patterns applied to the circuit. Hence, some
information about the typical input patterns applied to a circuit needs to be specified to estimate
its power dissipation. The most straightforward way to provide information about input patterns
is to actually provide a long input stream representing a typical usage pattern together with the
specification of the circuit. But it suffers from two drawbacks that the input traces can be
very large and cumbersome to manage and in many cases only incomplete information about
the environment may be available. Hence, instead, input signals are described by input signal
probabilities [89].
Definition 2.4. Signal probability: Let I(t), t E (-00, +00), be a stochastic process that takes
the values of logical 0 or logical 1, transitioning from one to the other at random times. The
signal probability of signal I(t) is given by
p(I)
= limT-too
+T
I(t)dt
(2.1)
-T
Definition 2.5. Signal Activity: The signal activity of a logic signal I(t) is given by
A(I)
nI(T)
= limT-too-y;-
(2.2)
where nI(t) is the number of transitions of I(t) in the time interval between -T/2 and +T /2.
More detail state signal probability (briefly state probability) calculation is discussed in [89].
Two methods are developed to calculate the state probability: Explicit methods and ADD based
methods.
21
Explicit methods use iterative method [44] or Gaussian elimination method [119] to compute
the state probabilities. The main limitation of this method is that it is not applicable to very
large size FSMs for which the state set is large enough to make even the storage of matrix P a
formidable task.
ADD based methods use Algebraic decision diagrams (ADDs) data structure to compute
the state probabilities to avoid manipulating the transition probability matrix. ADDs [13] are
"BDD-like" data structure.
The main difference is that an ADD has multiple terminals while a BDD only has two
terminals: 0 and 1. ADD based methods allow the manipulation of very large systems by
representing the transition probability matrix with an ADD.
The STG of a finite-state machine is implicitly represented by a BDD ( or, equivalently, by a
1/0-ADD) of its transition relation [30]. The transition relation is a Boolean function T{x, s, s').
The support of the transition relation consists of the input variables, the state variables and the
next state variables. T has value 1 when the STG of the machine has a transition from state
s to state s' with input x, zero otherwise. Similarly, the input probabilities can be represented
by an ADD. The ADD PI{x) is extracted from the array of input probabilities with the simple
formula PI{x) = II~(/PIi{Xi) where each PIi{Xi) is a single-node ADD with two leaves with
value Pi and 1 - Pi.
Given T and PI, the implicit representation of matrix P can be obtained by the following
symbolic formula:
P{x, s)
= PI{x) . LT{x, s, s')
(2.3)
s'
The ADD of P can be exponentially smaller than the traditional matrix representation.
Example 2.1. The STG of an FSM is given in Fig. 2.3{a) with input, output and conditional
transition probability at each edge. The transition relation T{x, s, s') is shown in Fig. 2.3{b)
in the tabular format. The ADD of T is shown in Fig. 2.3{c). The ADD of the conditional
input probability is shown in Fig. 2.3{ d). The result of the conditional transition probabilities
22
is shown in Fig. 2.3{e).

After computed conditional transition probabilities, the state probabilities can be computed
using the symbolic version of the power method[44]. The symbolic representation based on ADD
becomes useful when the STG and the truth table of T are unmanageably large.
s,x
s
1/0 (0.8)
1 0
1 0
(b)
(a)
(c)
(d)
(e)
Figure 2.3: Symbolic computation of the conditional transition probability matrix

Based on conditional state transition probability and the state probability, the state transition
probability can be calculated[89]. Because of no detail circuit structure available at this level,
switching activity is used to measure the power dissipation [88][77][15][12][48][117]. Given the
state encoding, each state is represented by binary code and Hamming distances can be calculated
for each state transition. Then, the switching activity of an FSM can be computed based on the
state transition probability and Hamming distances .
Information-theory-based approaches
23
Recently, information theory has been used quite effectively to estimate power at the RT (register transfer) level of design abstraction [75][62]. The RT level abstraction assumes that the
Boolean functionality of the circuit is known while the details of the implementation are unknown. Information-theoretic approaches depend on information-theoretic measures of activity
(for example, entropy) to obtain quick power estimate.
Entropy characterizes the randomness or uncertainty of a sequence of applied vectors and
thus is intuitively related to switching activity, that is , if the signal switching is high, it is likely
that the bit sequence is random, resulting in high entropy. Suppose the sequence contains t
distinct vectors and let Pi denote the occurrence probability of any vector v in the sequence.
Obviously,
2:;=1 Pi = 1. The entropy of the sequence is given by

t
h=
I:: Pi log2Pi
(2.4)
i=l
This equation is only an upper bound on the exact entropy, since the bits may be depended.
This upper bound expression is, however, the one that is used for power estimation purpose.
Furthermore, in [75], it has been shown that, under the temporal independence assumption, the
average switching activity of a bit is upper bounded by one-half of its entropy. Based on entropy,
the methods to estimate the low bound and high bound of an FSM switching activity have been
presented in [109][63] .
Power estimation included glitching
In the combinational section of the FSM, glitching power is inevitable. It has been observed that
this additional power dissipation is typically 20% of the total power, but can be as high as 200%
of the total power in some cases such as in a multiplier[123]. There are some publications on the
power estimation included glitching for FSMs [38][84][123]. Unit delay and general delay models
are proposed to estimate the glitching included power for circuits [42]. However, overestimation of
the power dissipation is still possible under unit delay model. Furthermore, multiple-option delay
models in a tool are not so convenient in practical applications. Monte-Carlo-Based approach
can estimate the exact glitching included power dissipation provided exact signal probabilities
24
and activities of primary inputs are known. However, accurate signal probability or activity
values for primary inputs may not often be available. Since power dissipation strongly depends
on the input signal properties, uncertainties in specifications of input signal properties make the
estimation process difficult[73]. Hence, more work needs to be done.
2.1.3
FSM power optimization
Different power optimization strategies have been applied at the different design levels. Here,
some important logic synthesis transformations will be discussed. Sequential logic optimization
methods work at two levels of abstraction, namely the State Transition Graph level and at the
logic-gate level. Several approaches have been developed in these levels as follows.
State assignment
State assignment and the resulting combinational logic synthesis have been conventionally targeted at reducing area and critical path delay [67][111][33]. For the optimization of power dissipation, these methods have to be modified to target a power cost function, namely, weighted
switching activity. State assignment significantly affects circuit power dissipation because different state assignments have different switching activities. Given an FSM, the state transition
probabilities between any two states can be obtained [15][48]. A state transition in an FSM
could be caused by single or multiple bit transitions. Therefore, the switching activity could be
minimized if states associated with state transitions that appear most frequently are assigned
codes that are close to each other. Several methods have been proposed for power-oriented state
assignment [88][77][15][48], which are mainly based on the minimal average Hamming distance.
Methods to encode State Transition Graphs to produce two-level and multilevel implementations
with minimal power are described in [117] and [108]. A method to re-encode logic-level sequential circuits to minimize power dissipation is presented in [43]. The problem of finding the state
assignment for the power optimization is computationally hard. There is no known method of
predicting the optimum assignment for the states though many algorithms have been proposed
[108] [77] [15] [48] [117].
Clock-gating
25
This scheme is based on the observation that during the operation of an FSM there are conditions
such that the next state and the output do not change (i.e., the machine is internally idle).
Therefore, clocking the FSM only wastes power in the combinational logic and in the registers in
this case. If the idle condition of the machine can be detected, the clock can be stopped until a
useful transition must be performed and the clocking is resumed. Fig. 2.4{a) is an input latched
FSM, which is different from the generic FSM structure in Fig. 2.1. Fig. 2.4{b) is a clock-gated
FSM structure. Fa, activation function, is to selectively stop the local clock of the FSM when
the FSM is internally idle. The block labeled "L" represents a latch, transparent when the CK is
low. The presence of a gated clock has a two-fold advantage. First, when the clock is stopped, no
power is consumed in the FSM combinational logic, because its inputs remain constant. Second,
no power is consumed in the registers and the gated clock line. In an arbitrary sequential
circuit, some parts of a circuit are not accessed in each clock cycle. A detailed discussion of this
technique is given in [16]. However, for those which have no internal idle states, the limitation
of this technique is obvious. To overcome this limitation, a new technique is presented in [90],
which can create idle states so that clock-gating technique can be used to save power. If simple
conditions that determine the inaction of particular registers can be determined, then power
reduction can be obtained by gating the clocks of these parts [17]. An extension version of this
technique is called precomputation technique which is presented in [2].
Combination
Logic
I FF
]N
STATE
IFF
]N
OUT
Combination
Logic
STATE
OUT
Fa
CK
CK
(a)
(b)
Figure 2.4: FSM structure model (a) single-clock, flip-flop based FSM model; (b) clock-gating
version
Partitioning
26
The fundamental intuition behind this technique is that a sequential circuit may be partitioned
into a set of small interacting blocks. During operation, only one block is active at any given time
and controls the input-output behavior. In the remaining blocks, the clock can be stopped and,
consequently, the total power consumption is reduced. FSM partitioning has been extensively
studied for several decades. Its theoretical foundations were laid down by Hartmanis and Stearns
[46] in the sixties. More recent work [41] reported experimental results on the implementation of
the partition procedures described in [46]. A different viewpoint on the problem was proposed
in recent years by Ashar, Devadas and Newton [10] who presented numerous algorithms for the
automatic partition of FSMs specified by an monolithic state transition graph (STG). However,
all these techniques are for minimum-area implementation. Recently, partitioning techniques
have been explored to reduce the power dissipation of FSMs. Chow et al.
[29] proposed a
low-power partition approach based on the relationship between state assignment and FSM
partition that produced very promising results. Dasgupta et al [32] proposed an approach which
is accomplished in two stages namely disjunctive partitioning and selective isolation encoding
while Monteiro and Benini suggested approaches to combine partitioning technique with clockgating[71][18]. The principle behind this technique is to partition an FSM into number smaller
interacting submachines such that only one submachine is active during any clock cycle. The
rest of circuit comprising of the other submachines is turned off, thus avoiding unnecessary
power-dissipating switching.
Retiming
=Ore!.
(a)
~
R
~
l
CL
(b)
Figure 2.5: Reducing the switching activity by inserting register
27
The transformation that repositions the registers of a design without modifying its external
behavior is called retiming. Monteiro et al. [70] have pointed out that register positions can also
affect power dissipation. Consider the simple example of a logic gate belonging to a synchronous
circuit as in Fig. 2.5(a), and call CL the capacitance load driven by the output node of AND
gate. In the case of CMOS technology, the power dissipated by AND gate is proportional to
the product of the switching activity of the output node of the gate a g and the output load
CL. Now consider the case in which a register R is connected to the output of AND gate.
Let CR be the input capacitance of the register, and let aR be the switching activity of the
register output [see Fig. 2.5(b)]. The total power dissipated by the new circuit is proportional to
agCR
+ aRCL < agCL if both a g and CL
are sufficiently high. Retiming repositions the flip-flops
in a synchronous sequential circuit so that the spurious transitions at the inputs to the flip-flops
can be filtered out by the clock as shown in Fig. 2.5. A retiming method that targets the power
dissipation of a sequential circuits is described in [70].
2.2
Low power design of flip-flops
It is found that although the power distribution of VLSI's differs from product to product a
clock system and its logic part consume almost the same power in various VLSI chips and the
clock system consumes 20%-45% of the total chip power. In the clock system power, 90%
is consumed by the flip-flops themselves and the branches of the clock distribution network
which directly drives the flip-flops [54]. Hence, low power design of flip-flops has attracted
many researchers [49][19][68][102][54]. Based on logic families, flip-flops can be grouped into pass
transistor based flip-flops[49][19], true single phase clocking (TSPC) based flip-flops, differential
designs and clocked CMOS (C 2 MOS) flip-flops.
Low power design of flip-flops have been carried out along the following lines .
Reducing the number of transistors used in flip-flops [49]
Reducing the number of transistors in flop-flops can reduce the internal nodes of signal transitions. Hence, it can not only save area but also implement global power savings.
28
Using double edges of clock signal to trigger flip-flops [110][49][19][102][58][40][1]

Compared to single edge-triggered flip-flops, the clock frequency for double edge-triggered flipflops can be reduced into half and hence clock system power is reduced. There are two kinds of
schemes to implement double edge-triggering. One is to use two latches or flip-flops to receive
input signals at both clock phases alternatively and use a multiplexer at the output section to
selectively output the stored signal[49][19][80]. This scheme often trades with the increasing
transistor number compared to the single edge-triggered flip-flops. The other is narrow pulsetriggered scheme. This kind of flip-flop is composed of two parts: a pulse generator and a
latch (or a flip-flop cell). The pulse generator usually consists of a series of inverters (3 or 4
inverters) while a latch (or a flip-flop cell) uses clock racing signals to generate a narrow pulse
corresponding to each of clock transition edges[102][72]. A good performance comparison for
variety double edge-triggered flip-flops is shown in [68].
Reducing clock signal swing to achieve power improvement[54]
This scheme is based on reducing clock signal swing to reduce the clock system power. However,
if there are clocked pMOS transistors in the circuit, it will make power supply complicated to
avoid static power dissipation increase.
Using clock gating techniques to deactivate clock signal so that power dissipation is reduced
[100][101]
Clock signal is the most active signal in a flip-flop if glitches are not taken into account. Clock
signal triggers a flip-flop and also triggers most of internal nodes. However, if input signal is not
changing or has low signal activity, the clock triggering is redundant. In these cases, if clock
signal can be deactivated, the power can be saved. The principle behind clock gating techniques
is based on this idea.
2.3
Power optimization of XOR gate based circuits
Any Boolean function can be expressed canonically based on AND and XOR operators using
Reed-Muller (RM) expansions. Because Reed-Muller realizations have several attractive advan-
29
tages especially for functions which do not produce efficient solutions using SOP techniques,
research on Reed-Muller logic has attracted more and more investigators [4][64][114][8][115].
However, most of these work is on area minimization. The basic approaches to minimize area
are:
Polarity optimization methods to find the best polarity with the least number of product
terms or literals.
To find the best polarity is computationally extensive in both space and time especially for large
functions. Traditionally exhausting search is only suitable for the smaller functions which are
less than 15 input variables[6][35][92]. However, progress has been made in [116], which can
solve large functions which have up to 25 input variables. Several heuristic methods have been
proposed which apply the simulated annealing [79] or genetic algorithm techniques[8][9].
Decomposition
Decomposition method is based on the concept of ~ majority cube [105]. The principle behind
this method is that an m-dimensional cube covers at least ~ x 2m on-set 7r-terms. The method
is further generalized to very large multiple output functions in [114].
Mixed polarity minimization
The product terms can be reduced with mixed polarity by combining the adjacent product terms
such as using XOR-link operation[97].
Compared to the area minimization, little work has been presented on power optimization
for XOR gate based circuits.
In [131]' a multiple input XOR gate is decomposed into a tree of two input XOR gates. The
aim is to search an optimized input signal combination which has minimum power dissipation.
In [74], FPRM functions are implemented into XOR trees and AND trees and power dissipation
is optimized with factorization and reduction rules. One limitation for these methods is that
they only optimize power under the specific function form or specific polarity.
2.4
30
Genetic algorithm
In this section, some basic operations of Genetic Algorithms (GAs) are introduced, which will
be further discussed in the next chapter.
2.4.1
Overview
First proposed by John Holland in 1975 [47] and then developed by his colleagues and students,
genetic algorithms (GAs) have been an attractive class of computational models that mimic
natural evolution to solve problems in a wide variety of domains. GA emulates biological evolutionary theories to solve optimization problems and composes of a set of individual elements
(the population) and a set of biologically inspired operators defined over the population itself.
According to evolutionary theories, only the most suited elements in a population are likely to
survive and generate offsprings, thus transmitting their biological heredity to new generations.
The basic procedure is to create a population (breeding pool) of potential solutions to a problem. These solutions are encoded as "chromosomes" (data representation of the solution), and
each chromosome is subjected to an evaluation function which assigns "fitness" depending upon
how well the solution it encodes solves the problem at hand. Existing solutions are recombined
by a process called crossover or breeding. The rational for this is that good solutions will contain
good building blocks, rearrangement of which may produce even better solutions. Further a mutation process makes random changes in a few randomly selected chromosomes. This prevents
premature convergence by maintaining the diversity of the population. GA operates through a
simple cycle of stages:
Creation of a "population" of chromosomes
Evaluation of each chromosome
Selection of "best" chromosomes
Genetic manipulation to create new population of chromosomes
Fig 2.6 shows these four stages using GA.
31
OHAPTER 2. OONVENTIONAL LOW POWER DESIGN TEOHNIQUES
2.4.2
Representation
Fundamental to the GA structure is the encoding mechanism for representing the optimization
problem's variables. The encoding mechanism depends on the nature of the problem variables.
In each case the encoding mechanism should map each solution to a unique binary string.
Example 2.2. Consider optimizing the function f(x)
= x 2 where the continuous variable x is
defined in a range from [0,2] with an accuracy of two decimal places after the decimal point.
The variable can be encoded by a binary code. The mapping from a binary code into a real
variable value within the range [0, 2] can be implemented as in equation 2.5.
2x
(2.5)
x = 28 _ 1
Here, x' is binary code < b7b6'" b1bo
> and 2 is the length of the domain. The reason to
choose 8 bit binary code is to meet the accuracy because
128
= 27 < 200 < 28 = 256

Population
(chromosomes)
Decoded
chromosomes
Offsprings
New generation
Parents
Genetic
operators
Evaluation
(fitness)
Manipulation
Mates
Figure 2.6: GA cycle
32
For example, a chromosome x' =(01101100) represents 0.85, because
x' = (01101100h = (108ho
and
108
x = 2 - = 0.85
255
The chromosomes (00000000) and (11111111) stand for the boundaries of the domain, 0 and
2.0, respectively.
2.4.3
Initial population
A population of chromosomes needs to be initialized. Any of possible 8-bit binary codes could be
a chromosome in the above case, which is initialized randomly. The population size (the number
of chromosomes) is set depending on the application.
2.4.4
Evaluation
An evaluation function returns a measurement of the worth of any chromosome in the population.
Evaluation function Eva(x') is equivalent to the function f(x).
Eva(x') = f(x)
For example, x' = (011011000) corresponds to x = 216.
Eva(01101100)
= f(0.85) = 0.72
Usually, Eva(x') or f(x) needs to be transfered to fitness function fit(x'):
33
Jit(x') = h(Eva(x')) = h(f(x))
The fitness stands for a measure of how good the chromosome is.
2.4.5
Parent selection
Selection models nature's survival-of-the-fittest mechanism. Fitter solutions survive while weaker
ones perish. There are many ways to do this. The most popular parent selection scheme is the
roulette wheel parent selection. It works by allocating pie-shaped slices on a roulette wheel to
population members, with each slice proportional to the population member's fitness. Selected
parent can then be viewed as a spin of the wheel, with the winning population member being
the one in whose slice the roulette spinner ends up.
2.4.6
Crossover
Crossover recombines the genetic material in two parent chromosomes to make two children. The
simple crossover is one point crossover, which occurs when parts of two parent chromosomes are
swapped after a randomly selected point, creating two children. Fig. 2.7 shows an example of
the above application of one point crossover supposing that the cut point is selected after the
6th bit (usually called gene).
Child 1: 011000 I 10
Parent 1: 011000 I OJ
---l>-
Parent 2: 110110 I 10
Child 2: 110110 I 01
Figure 2.7: Crossover operator
2.4.7
Mutation
Mutation is a secondary operator with the role of restoring lost genetic material. It also reduces
the possibility of early convergence on a local optimum solution. Mutation of a bit involves
flipping it from 0 to 1 or vice verse. For example, for child 1: 01100010, if the sixth gene
34
is selected for a mutation, it would be flipped from 0 to 1 and result in a new chromosome
01100110.
2.4.8
Elitism
After crossover and mutation, the best members of the population may fail to produce good
offsprings in the next generation. To restore the possible loss, the best members of each generation
are copied into the succeeding one.
Chapter 3
State assignment for area and power

optimization
In this chapter, a genetic algorithm (GA) based state assignment targeting area and power
optimization is developed.
3.1
Finite state machines (FSMs)
A finite state machine (FSM) is represented by a set of states and a set of their associated
transitions.
An FSM can be represented by a graph or, equivalently, by a table. The two representations
are called state transition graph (STG) and state transition table (STT), respectively. The states
of an STG are labeled with the unique symbolic state names. The edges are labeled with the
input and output values. The state table is simply the list of edges of the STG. Take DK27 in
1991 MCNC benchmarks [129] for example as shown in Fig. 3.1.
FSMs can be incompletely specified. An incompletely specified FSM is one where 'T/(x, s) and
lor .\(x, s) are incompletely specified Boolean functions. To synthesize an FSM, the symbolic
state names must be encoded as unique binary codes. The binary codes assigned to the symbolic
states determine the circuit's combinational logic. Then state assignment is defined in Definition
3.1.
35
CHAPTER 3. STATE ASSIGNMENT FOR AREA AND POWER OPTIMIZATION
36
Definition 3.1. 8tate assignment is a mapping from the set of states of an F8M to the set of
binary codes.
It is well known that an F8M's state assignment can significantly affect the quality of syn-
thesized circuits. Much of state assignment research has been concentrated on reducing the
circuit area.
NOVA[111] makes state assignments which target minimal-area two-level logic
while MU8TANG[33], JEDI[57] and MU8E[36] target multilevel-logic implementation. Recently,

several researchers have focused on low-power designs using state assignment that reduces the
average switching frequency of the states [88][12][108][77][15][48][117]. The 8yclop [88] method
considered conditional state transition probabilities as weight coefficients in the cost function,
while Hong [48] and Wang [117] methods exploited the total state transition probabilities. The
shortcoming of the above approaches is that they minimize the switching activity on the present
state bits without any consideration of the area in the combinational section of F8Ms. As a result,
the area overhead is high compared to area-oriented state-of-the-art tools. To consider the area
constrain in combinational section, Olson [77] and POW3 [15] introduced a convex combination
(a)
In
P8
N8
0
0
0
0
0
0
0
1
1
1
1
1
1
1
81
82
83
84
85
86
87
86
85
84
87
81
52
53
86
85
85
86
81
81
85
82
82
56
56
84
53
57
00
00
00
00
10
01
00
01
10
10
10
00
00
00
(b)
Figure 3.1: 8tate transition graph and state table of DK27 (a) STGj (b) 8TT
37
of the switching activity and the area of the combinational logic as cost function. Wang [113]
proposed a matching-based state assignment algorithm to minimize area and state transitions
simultaneously. However, all of those results show that minimization of switching activity has to
trade with area penalty. As we know, area overhead will result in many disadvantages such as
cost and reliability of circuits. Further, high area overhead will in turn offset the power reduction
in state registers.
The problem of finding the state assignment for the minimization of power consumption and
area is computationally hard. The two primary techniques used to solve the problem are the
greedy search and the simulated annealing. However, the search space appears to be too large
with many local minima for these schemes to find the global minimum. The genetic algorithm
(GA) technique has been successfully applied to a variety of computationally difficult problems
which have a large search space. It has been shown that it can produce good results in reasonable
computation time. A recent investigation showed that GA can find better assignments than
commercial products for area minimization [5]. Olson [77] employed a genetic local search to
perform a local optimization of FSMs and got encouraging results. In this Chapter, a new scheme
is proposed and genetic algorithms (GAs) are employed to optimize both switching activities and
area without the need to carry out an exhaustive search.
The remainder of the chapter is organized as follows. The terminology used here is defined
in Section 2.1.1. Section 3.1 introduces the FSM calculation while Section 3.3 defines the cost
function. The state assignment algorithms are described in Section 3.4. Experiment results and
conclusions are given in Sections 3.5 and 3.6, respectively.
3.2
Terminology and parameter calculation of FSMs
Power dissipation is a strongly pattern-dependent cost function. Here, input signals are described
by input signal probabilities.
Given a circuit, suppose the input signal probability, Pi, is known. However, in an FSM, the
probability of a state transition depends not only on the inputs but also on the state information.
Considering a transition between two state
Si
and
Sj.
If state Si is unreachable, the machine will
38
never perform the transition because it will never be in state Si. Similarly if the probability of
being in state Si is very low, a transition from state Si to state Sj is very unlikely. A parameter
to describe the probability of a transition is called conditional probability.
Definition 3.2. Given a set of inputs {Ioh ... In-I}, Conditional State Transition Probability
(CSTP) Pij associated with a transition from Si to Sj (briefly tSij) is the ratio of the number of
input minterms causing such transition to the total number of valid input minterms at state Si.
i, j E {O, 1,2"", n - 1} and n is the number of states.

If the input probability is not specified, a default input probability of 0.5 is used. In order to
simplify the calculation of the conditional probability, assumption of Markov chain is employed
[14].
Definition 3.3. A Markov chain is a representation of a finite-state Markov process, a stochastic
model where the transition probability distributions at any time depend only on the present state
and not on how the process has reached in that state. The Markov chain model for the STG is
a directed graph isomorphic to the STG and with weighted edges.
Symbolically the parameter can be expressed as:
Pij
Here, i, j
= 0, 1,2, ... , n
= Prob(Next = sjlPresent = Si) =
Nij
2:k Nik
(3.1)
- 1, Nij is the number of transitions tSij from Si to Sj while 2:k Nik
is all transitions that begin in state Si.

Given an FSM, calculation of the conditional state transition probability Pij is straightforward
assuming uncorrelated and equiprobable inputs for simplicity [15]. The following example shows
how to calculate Pij given input signal probabilities
Example 3.1. Consider the FSM shown in Fig. 3.1(a) with one input, In, and two outputs.
Assume that the input probabilities are Prob(In = 1) = 0.5. The conditional transition probabilities are labelled on the edges of the Markov chain. For instance, consider the transition tS14.
Its CSTP is PI4
= Prob(In = 1) = 0.5.
The detail is shown in Fig. 3.2.
39
This process is simple if the inputs of the FSM are completely specified as in Example 3.l.
However, if the inputs are incompletely specified, the cases of overlapping inputs and impossible
inputs need to be dealt with, which is shown in the following example.
Example 3.2. Table 3.1 shows some lines of an incompletely specified FSM tav, which is one
of MCNC benchmark circuits and given in the kiss file.

The first four lines tell the number of inputs, outputs, product terms and states, respectively.
The machine has four inputs, four outputs, 49 product terms and four states. From the fifth
line, each row consists of four sections, which give input encoding, present state, next state and
output encoding. For the transition
tS01,
the maximum number of the transition is sixteen. It
can be seen that the input cases {0111, 1011, 1101, 1110, 1111} have been over calculated for
the incompletely specified inputs. Thereby, the overlapping inputs need to be checked, which
can be expressed into
Il nIr
:f: e
Here, I l , Ir are input vectors resulting in the same transition

of the circuit description file and
e is an empty set.
(3.2)
tSij,
:f: r
are the line number
Algorithm 3.1 outlines the procedure used
114=112
Figure 3.2: Conditional state transition probability of DK27
40
to calculate Pij.
Algorithm 3.1. Procedure for finding GSTP
Find_ GSTP(8i,8j)
{
Pij = L:P(Il);
for (l = 4; l < MaximumLineNumber - 4; l + +) / /MaximumLineNumber
stands for the maximum line number of the circuit description file
for(r = l + 1; r < MaximumLinNum - 4; r + +)
{
if Il n Il- 1 =1= I>
{
Pij = Pij - L:IEI/ nI/-l p(I);
}
}
Table 3.1: Tav's kiss format
.i4
.04
.p 49
.s 4
1000
0100
0010
0001
0000
111-11-1
-11-1-1
-11
1000
80
81
80
81
80
81
80
81
80
81
80
81
80
81
80
81
80
81
80
81
80
81
81
82
1000
0100
0010
0001
0000
0000
0000
0000 I
0000 I
0000
0000
1000
41
The CSTP is external input information. It does not depend on the structure of the Markov
chain. For an FSM, this information can be assumed to be known. Some researchers used
this information as a rough approximation to the transition probabilities [88]. However, this
probability does not utilize the information of state probabilities, where the state probability,
Pi, represents the probability that the machine is in a given state Si.
Definition 3.4. State probability Pi of a state Si, which is defined as the probability that the
state is visited in an arbitrarily long random sequence, can be obtained by solving the corresponding Chapman-Kolmogorov equations and the normality condition equation in equation
3.3.
",i=n-l n. L.ii=O
r~-
. _
...
P~ -
2::
j=n-l
j =o
(3.3)
PJPJ~
By solving the above set of linear equations, the state probability Pi can be obtained, which
is shown on the edges of the Markov chain as in Fig. 3.3.
P2 =4/21
P3 =2/21
Figure 3.3: State probability labelled beside its state node and state transition probability labelled on each edge of DK27
42
The well-known Gaussian elimination method is used to solve the above equations [119] . The
core of the method is to convert a matrix into an upper triangular form and solve for Pi in
AP
= b using the back-substitution method. Here, A is an upper triangular matrix, P is state
probability vector and b is a constant vector based on equations 3.3. Algorithms 3.2 and 3.3
outline the procedures.
Definition 3.5. An upper triangular matrix is one in which all elements below the main diagonal
line are zero as follows
aOO
aOl
a02
aO(n-l)
o
o
au
a12
al(n-l)
a22
a2(n-l)
000
a(n-l)(n-l)
Algorithm 3.2. Convert a matrix into an upper triangular matrix
UpperTriangular{s) {
for (i; ; ;) {
== 0) {
if (A(i, i)
sort(A); }llsort the matrix so that the diagonal element is not zero
else
pivot = A[i][i];
for (j = i
+ 1;
;){
mult = A[j][i]lpivot;
A[j][i] = 0;
for(k
= i + 1; ; ) {
A[j][k] = A[j][k] - mult * A[i][k];}
b[j] = b[j] - mult * b[k];}}}
43
Algorithm 3.3. Back substitution

=====================================================================================================
Backsubstitution() (
for(i; ; ;) {
forU
+ i + 1;
sum
b[i]
;)
= sum + A[i]Li] * b[j];
= (b[i]- sum)/A[i][i];}}//The
answer is returned by b
State transition probability can be defined as follows.
Definition 3.6. State transition probability (STP) tpij between two states Si to Sj occurs in an
arbitrarily long sequence and is given by
tPij = PiPij
(3.4)
The STPs of DK27 are shown in Fig. 3.3.
Definition 3.7. The switching activity of the state bit lines depends on the state encoding and
the state transition probabilities. Average switching activity of an FSM can be calculated as
follows:
i=n-lj=n-l
SA
L L
i=O
tPij x HD(enc(sd, enc(sj))
(3.5)
j=O
where enc(sd is encoding of state Si and HD(enc(sd, enc(sj)) is the Hamming distance between two encodings, enc(si) and enc(sj). From equation 3.5, for an FSM with given input signal
probabilities, {tPij} is fixed while H D( enc( Si), enc( Sj)) varies with different state assignments.
Thereby, SA varies with different state assignments. This is the basic principle to optimize power
dissipation using state assignment.
3.3
44
Cost functions
The implementation of an FSM consists of two parts: a combinational logic section and a register
section. Both sections contribute to the power dissipation of an FSM. Traditionally, switching
activity is used to measure the power dissipation of a circuit. Some researchers [48][117] suggested
the use of switching activity as cost function. The approach implements the low power dissipation
by minimizing equation 3.5. However, this only minimizes the switching activity on present state
bit lines of the machine and does not consider the structure of the combinational section of the
final synthesized FSM, which may lead to non-optimal area implementation and result in power
overhead in the combinational section. Hence, to obtain low power dissipation in the final circuit,
area should be taken into account.
In [15], the cost function is linear composition of the weighted area and the switching activity.
Fan-in- or fanout-oriented method [33] based weighted area was used to indicate the desired
Hamming distance between the state codes. The higher the weight is, the smaller the Hamming
distance will be. A parameter a
1 was introduced, specifying the relative importance of
switching activity with respect to area constraints. It was defined as
i=n-lj=n-l
Cost = (1 - a)SA + a L L
i=O
Here,
Wij
is weighted area between
Si
and
Wij
(3.6)
j=O
Sj.
In [77], literal based area was employed and the cost function is normalized to 1000.
Cost
= b.literals + f3
x SA
(3.7)
There, f3 was taken as 100.

The cost function is one of very important factors to guide circuit optimization. In this work,
two kinds of cost functions will be used and described as follows.
In Section 3.5, area and SA versus a and f3 will be studied. However, the number of cubes
45
after the logic minimization of a machine is used instead of weighted area and literals used in
equations 3.6 and 3.7. It will be shown that if the cost function is normalized to 1, SA is very
sensitive to change in a while area is less sensitive (See Fig. 3.4(a)). Further, ifthe cost function
is normalized to 1000, area is very sensitive to change in {3 while SA is less sensitive (See Fig.
3.4(b)). To influence both SA and area, a factor of 10 is used here and the first cost function is
defined as in equation 3.8 [124].
Costa
= noOfCubes + SA
(3.8)
Here, noO f Cubes is the number of cubes after the logic minimization of a machine.
On the other hand, for CMOS circuits, the dominant source of power dissipation is the
charging and discharging of the node capacitance and is given by:
1
Paver
= "2 Vldfclk
k=m
2: CkSAk
(3.9)
k=l
Where Paver is the average power dissipation of the circuit, Vdd is the supply voltage, Ck is the
capacitive load at the output of gate k, fclk is the clock frequency, SA k is the switching activity
of gate k and m is the number of gates in a circuit and k E {1, 2" . m}. All of the parameters
in the above equation can be determined from technology or circuit layout information except
SA k and Ck of the circuit which depend on the synthesis technology. Hence, equation 3.9 can
be rewritten as in equation 3.10.
k=m
Paver = K
2: CkSAk =
k=l
k=m
KSA
2: Ck
(3.10)
k=l
Here, K =~ Vldfelk while SA is the average switching activity of a circuit and defined as
SA
1..
",k=m SA k
m L.Jk=l
Hence, we have Paver ex Area x SA, where Area is estimate of the circuit area that is
46
representative of the capacitance 2.:~~~ Ck [75].

Given an STT of an FSM and a specific state encoding, SA can be calculated by equation
3.5. Under the circumstance of employing the minimal encoding length, the number of memory
elements used in a specific FSM are constant so that their corresponding area is fixed. Hence,
only the area of the combinational section varies with the state assignment. The area of the
combinational section could be measured by the number of cubes. Therefore, the second cost
function is defined as in equation 3.11.
Costb
3.4
= noO fCubes
x SA
(3.11)
State assignment using Genetic algorithm
Genetic Algorithms (GAs) are based upon evolutionary adaptation in natural systems and attempt to generate useful solutions to a given problem by the application of the "survival of the
fittest" principal. The basic idea is to create a population (breeding pool) of potential solutions
to a problem. These solutions are encoded as "chromosomes", and each chromosome is subjected
to an evaluation function which assigns ''fitness'' depending upon how well the solution it encodes
solves the problem at hand. Existing solutions are recombined by a process called crossover or
breeding. Further, a mutation process makes random changes in a few randomly selected genes.
This prevents premature convergence by maintaining the diversity of the population. The best
members of the population may fail to produce offspring in the next generation. The elitist
strategy rectifiers this potential loss by copying the best members of each generation into the
succeeding one.
3.4.1
Solution representation
An FSM with n states requires a minimum of s state variables for the assignment where s
rlog2m1 and symbol
rl
stands for taking the upper bound integer of log2m. The chromosome
representation is a string of decimals.
Take benchmark 'DK27' for example.
It is a seven
CHAPTER 3, STATE ASSIGNMENT FOR AREA AND POWER OPTIMIZATION

state machine and the states are named
7 is one possible chromosome, Then,
S1, S2, S3, S4, S5, S6, S7'
S1
47
Assignment 6, 2, 1, 4, 3, 5,
is assigned the binary code 110,
S2
is assigned 010,
etc, The population of solutions consists of state assignments of an FSM, They will be stored
in a two dimensional array population[i][j], where, (i=0,1,2, "" populationSize-1, j=0,l,2, ""
numberOfStates-1) , The breeding pool of chromosomes is initially created by InitialPopulation
orandomly, The sketch of InitialPopulation 0 is as shown in Algorithm 3.4:

Algorithm 3.4. Initializing population
InitialPopulation (int numberOfBits, int numberOfStates )
(
int maxInt, randNum, randRemainj
int doubleEentryFlag
maxInt
OJ
1 < < numberOfBitsj
for(i=Oji<populationSize;i++ )
for(j=O;j<numberOfStates;j++ )
population[ilbJ
maxInt +1;
srand (seed Value};

for(i=O;i <populationSize;i+ +)
for(j=O;j<numberOfStates;j++ )
{
do
(
randNum
randRemain
rand{}j
randNum%power;
for(k=O;k<numberOfStates;k++ )
if(randRemain==population[i}{kj)
doubleEntryFlag = 1;
} while (doubleEntryFlag==1);

population[i}[j}
48
= randRemain;
}
}
3.4.2
Evaluation
Based on the above assignment, the cost can be calculated according to equations 3.8 & 3.11.
To a specific state assignment for an FSM, let the maximum state transition probability
tp
= max(tPij)
(3.12)
and maximum Hamming distance between two state codes
maxHD
= max(HD(enc(Si), (Sj))) = lengthOfStateCode
(3.13)
The maximum SA of the FSM can be calculated as:
i=n-lj=n-l
maxSA
2: 2:
i=O
tp x maxHD
(3.14)
j=O
The area can be estimated by the number of cubes after state assignment. maxCubes takes
the product terms from the STT. Then the maximum costs will be:
maxCosta
maxSA + maxCubes and maxCostb
max SA x maxCubes.
Hence, to a specific state assignment, the fitness can be defined as:
fitnessmode = (maxCostmode - costmode)/maxCmode
(3.15)
49
Here, mode E {a, b}, which corresponds to two cost models in equations 3.8 & 3.11, respectively. From equation 3.15, lower cost will result in higher fitness.
3.4.3
Crossover
Parent selection strategy follows the one proposed in [9]. Crossover is a primary method of
perturbations in GA, which generates better solutions by exchanging the information contained
in the present solutions. In this application, a position-based crossover (PBX) is employed
[130]. This was modified, however, so that invalid offsprings are avoided and efficient crossover
is reached. This is outlined as follows.
Randomly select a number oflocations named location-indicator (LID) in binary codes whose
length is equal to the number of states in the FSM. Where 1s appear in the LID, copy the states
from Parent 1 to Child 1. Where there are Os in the LID, copy those corresponding states from
Parent 2 provided that they do not exist in Child 1. If the state from Parent 2 is already in
Child 1, the position is filled by the first unassigned state from Parent 1 but checks are made
to avoid duplicating the parent. Continuing this way, Child 1 is obtained. With the same LID,
interchange Parent 2 and Parent 1 and repeat the process to produce Child 2. This is illustrated
for a seven state machine.
Step 1. Randomly generate a seven-bit LID: 1 0 1 1 0 1 0
Step 2. Select two parents by the roulette wheel approach:
Parent 1: 6 2 1 4 3 5 7
Parent 2: 1 0 5 3 2 4 6
Step 3. Generate Child 1:
a. Where 1s appear in the LID, copy the states from Parent 1 to Child 1 and delete those
copied states in Parent1.
Parent 1: - 2 - - 3 - 7; Child 1: 6 - 1 4 - 5 -;
b. Where there are Os in the LID, copy the states from Parent 2 to Child 1 provided that
they do not exist in Child 1 and delete the copied states in Parent 2. There are states 0, 2 and
6 from Parent 2 corresponding to Os in the LID. However, state 6 is already in Child 1. Hence,
only states 0 and 2 in Parent 2 are copied to Child 1.
50
Parent 2: 1 - 5 3 - 4 6; Child 1: 6 0 1 4 2 5 -;
c. The unfilled position in Child 1 is filled by the first unassigned state from Parent 1, which
is state 3. Then state 3 is copied to Child 1.
Parent 1: - 2 - - - - 7; Child 1: 6 0 1 4 2 5 3;
Step 4: Swap Parent 1 and Parent 2 and generate Child 2 following the same procedure.
3.4.4
Mutation
Mutation in normal GA just flips a selected bit from 0 to 1 or vice verse. We mutate CHROMOSOMES state number one by one with a mutation rate 6%. The procedure is as follows:
For example, for a 7-state FSM, if the chromosome is 6 2 1 4 3 5 7, the following procedure
will be executed. At first, the program will randomly generate a number, namely randomNum1,
and check the first state 6, and if randomNum1 < 6%, produce two different random numbers
randomNum2 and randomNum3, for instance, randomNum2
2 and randomNum3
4,
then, the state numbers in the third bit and fifth bit of the chromosome will be swapped and the
chromosome would become 6 2341 57. Then it will check the second state, if randomNum1 >
6%, no exchange takes place in the chromosome during this step. The program will check the
generations one by one until the last generation. The detail is shown in Algorithm 3.5.
Algorithm 3.5. Mutation operator
====================================================================================================
Mutation (numberOfStates)
(
for(i=O;i<numberOfStates;i++ )
(
randomNuml
= rand()%100;
if(randomNuml < MUTATION_RATE)
(
randomNum2 = rand()%(numberOfStates);
do
51
(
randomNum3
rand() %numberOfStates;
}while (randomNum2 !=randomNum3);

swap (population[randomNumlJ[randomNum2j, population[randomNumlJ[randomNum3J);
}
}
}
3.4.5
Elitism
The best member of the population may fail to produce offspring in the next generation. Hence, it
will be kept in the array eletistM ember[i] and stored in the array population[populationSize/2l][i] of the next generation, where i = 0, 1,2, ... , numberOfstates - 1.
3.4.6
Outline of the algorithm
The outline of the algorithm is illustrated in Algorithm 3.6.
Algorithm 3.6. Outline of the algorithm for state assignment

====================================================================================================
Step 1. Read the benchmark

Step 2. Generate an initial population-brooding pool
Step 3. Calculate state transition possibilities tpij
Step
4.
Assign states and create a Berkeley standard PLA file for each chromosome
Step 5 Calculate the number of cubes by ESPRESSO minimization for each chromosome
Step 6. Calculate the fitness of each chromosome
Step 7. Select parents by the roulette wheel approach.
Step 8. Perform crossover to generate offsprings
Step 9. Mutate chromosomes
Step 10. Assign states and create a Berkeley standard PLA file for each chromosome
52
Step 11. Calculate the number of cubes by ESPRESSO minimization

Step 12. Calculate the fitness of each chromosome
Step 13. Generations
Step
14.
generations
+1
If generations'5:.max_ numberGenerations, goto Step 7.
Step 15. Output the results:

(1) The half chromosomes with higher fitness
(2) The best state assignment, its switching activity and the number of cubes
3.5
The above algorithm is implemented in 0 and applied to MONO benchmark circuits. Gaussian
elimination method [119] is used to find the total transition probabilities according to the STT
of an FSM. ESPRESSO is used to minimize the circuit after state assignment and obtain the
cubes. This is done by generating a Berkeley standard PLA file for each chromosome and passing
it to ESPRESSO for minimization. The product terms from this minimization determine the
number of cubes for that assignment. Switching activity is calculated using equation 3.5. TWo
sets of experiments have been conducted.
First, we want to find out how the combination of area and SA, two cost functions in equations
3.6 and 3.7, affect the optimization quality. In order to compare them, we use noO fCubes instead
of w in equation 3.6 and filiterals in equation 3.7, respectively. Benchmark circuits 'ex4' and 'cse'
are chosen as study cases. The reason for choosing them is that 'ex4' has 14 states and is an
incompletely specified machine while 'cse' has 16 states and is a completely specified machine.
Both are medium size FSMs. Figs. 3.4(a) and (b) show that area and SA vary with a and (3,
respectively. The plots show:
The area ratio Anoo fCubes I AaverageO fCubes
The SA ratio SAl SAaverage
From Fig. 3.4(a), we notice that area varies much less than SA when a varies from 0.1 to 0.9.
From Fig. 3.4(b), SA varies much less than area when (3 varies from 100 to 900. It makes sense
53
that equation 3.6 based cost function does not have too much influence on area minimization
while equation 3.7 based cost function has little influence on SA minimization.
1.6
____ a re a_c;!; e
1.S
_____ SA _c;!; e
1.~
1.3
1.2
j\
---.- a re a_ e x 1
---ff"-
SA _ex 1
1 .1
0.9
0.8
o .7
0.6
o .1
0 .2
0 .3
0 .~
0 .S
0 .6
0 .7
0 .8
0 .9
a h'a
(a)
1.2
1.1
0.9
-t--area_cH
___ SA_C;!;e
0.8
----t- a re a_e xl
0.1
______ SA _ (!X 1
0.6
100
300
SOD
100
900
be ta
(b)
Figure 3.4: Relationship between cost function and optimization quality (a) Area and SA versus
a in equation 3.6; (b) Area and SA versus (3 in equation 3.7
Second, fourteen MCNC benchmark circuits, whose number of states are between 10 and
54
48, are chosen to test the proposed methods. Then an area-oriented state assignment program,
NOVA [111], is run on the same benchmarks, following the same procedure to get the switching
activity and the number of cubes. Tables 3.2 summarizes the test results. The results given by
Hong et al and Wang et al are listed in the same table for comparison. In Table 3.2, cubes a
and BAa are obtained using costa in equation 3.8 while cubesb and BAb are obtained from costb
in equation 3.11. Compared to NOVA, Hong and Wang gave 47.5% and 42.3% reduction in
switching activity while 10.3% and 12.6% extra product terms are used, respectively. Using
costa, GA gives a 32.2% reduction in switching activity and 4.4% reduction of product terms.
Among the 14 circuits, our method gives better area results in 11 cases compared to NOVA
and in 10 cases compared to Hong's and Wang's. Using costb, GA gives a 43.8% reduction in
switching activity but needs 5.7% extra product terms. The area penalty, however, is much
smaller than Hong's and Wang's. Table 3.3 gives the two set of codes, namely a and b, for the
best state assignment by GA, which are based on costa and costb, respectively.
The above results are obtained using the following parameters: population size = 70, crossover
rate = 60%, mutation rate
= 6%, and maximum number of generations = 200. These parameters
were determined after testing various population sizes and different crossover and mutation rates.
3.6
Summary
A genetic algorithm for finding good state assignments targeting minimization of power and area
for finite state machines has been developed and implemented in C. Tests as to the effectiveness
of this approach to the problem are conducted by comparison of performance against the state-ofthe-art commercially available software and some published results when operating upon MCNC
FSM benchmark circuits. Two options are available for optimizing FSMs targeting area or power
dissipation. The results shows that significant saving in power and/or area can be achieved.
@
~
"\j
~
cr"
CTl
I:.I.:l
benchmarks
Name
States
bbara
10
bbsse
16
16
cse
downfile
24
keyb
19
modulo12
12
planet
48
sl
20
sand
32
styr
30
ex1
20
ex4
14
opus
10
train 11
11
Total
aver.%red.
GA
cubes a
22
27
43
36
46
12
86
66
89
88
52
14
15
10
606
4.4
BAa
0.317
0.783
0.355
1.600
0.674
0.583
2.424
1.480
0.765
0.943
0.842
0.421
0.556
0.339
12.082
32.2
cubeSb
24
28
44
47
48
11
92
76
101
98
59
16
16
10
670
-5.7
BAb
0.279
0.788
0.252
1.438
0.573
0.583
1.682
1.184
0.664
0.586
0.750
0.467
0.417
0.339
10.002
43.8
Wang
cubes
26
31
48
45
58
12
103
91
109
99
47
18
17
10
714
-12.6
[117]
SA
0.279
0.776
0.239
1.125
0.556
0.500
0.984
1.175
0.610
0.553
1.135
0.957
0.712
0.714
10.315
42.3
Hong
cubes
26
31
50
40
52
12
101
85
110
101
49
16
17
9
699
-10.3
[48]
SA
0.295
0.856
0.292
1.083
0.647
0.583
1.153
1.131
0.604
0.578
0.755
0.495
0.524
0.360
9.356
47.5
NOVA [111]
SA
cubes
24
0.495
1.500
30
46
0.604
28
1.750
1.466
48
12
1.000
2.831
87
1.698
80
1.085
97
94
1.278
44
1.358
19
1.316
0.812
16
9
0.619
17.812
634
0
0
tIj
~
'0
CTl
I-j
S
CTl
e.
I-j
CTl
S..:l
CJ:l
~
~
~
CJ:l
~f..3
~
.,....
~.....
m
m
::r"
::J
Q'q
'0
0
~
~
I-j
"\j
[?;'
..,
..,
S
I-j
CTl
'0
I-j
CTl
S
CTl
::J
.,....
~~
0
"\j
bl
~
~
bl
0
<
56
Table 3.3: Best state assignment codes by GA: Assignments a & b are based on costa & costb
respectively
a: 8, 0, 2, 7, 4, 6, 15, 14, 10, 12 ; b: 8, 0, 4, 6, 2, 3, 7, 14, 12, 10
bbsse
a: 8, 2, 6, 14, 3, 10, 15, 13, 9, 12, 11, 0, 4, 1, 7, 5
b: 4, 14, 10, 3, 6, 2, 9, 1, 8, 11, 0, 12, 13, 15, 5, 7
a: 0, 1, 13, 5, 15, 6, 2, 3, 9, 7, 11, 10, 8, 14,4, 12
cse
b: 0, 4, 12, 6, 14, 11, 8, 10, 1, 2, 3, 7, 5, 15, 9, 13
donfile
a: 2, 3, 0, 4, 1, 5, 6, 7, 10, 11, 8, 9, 13, 12, 14, 15, 21, 23, 17, 16, 18, 19,
25, 27
b: 11, 5, 12, 28, 1, 8, 15, 7, 14, 26, 31, 29, 13, 20, 6, 30, 0, 4, 3, 24, 23,
21, 16, 22
keyb
a: 0,4, 16,8,21, 12,20, 19,7, 17, 14, 1, 15, 23, 3, 9, 28, 10, 13
b: 0, 4, 2, 8, 31, 10, 27, 25, 24, 17, 15, 3, 13, 29, 9, 28, 16, 21, 5
modulo12 a: 0, 9, 13, 15, 11, 3, 7, 5, 4, 6, 14, 12; b: 9, 8, 10, 11, 15, 7, 0, 2, 3, 1,
5,13
planet
a: 26, 11, 7, 0, 4, 10, 18, 24, 22, 15, 48, 43, 59, 14, 9, 12, 46, 44, 8, 63,
51,6, 5, 13, 47, 34, 23, 52, 2, 1, 58, 45, 61, 37, 25, 35, 42, 32, 31, 60, 39,
56, 50, 30, 57, 49, 62, 21
b: 50, 36, 52, 20, 16, 2, 14, 6, 54, 53, 46, 49, 40, 51, 9, 11, 3, 19, 25, 56,
57, 35, 7, 32, 24, 28, 29, 21, 8, 33, 12, 45, 5, 37, 13, 39, 61, 63, 31, 23,
15, 62, 58, 18, 22, 4, 48, 26
sl
a: 20, 18, 10, 26, 5, 27, 16, 24, 0, 8, 30, 25, 29, 13, 22, 1, 19, 9, 4, 21
b: 13, 5, 28, 29, 10, 31, 1, 21, 0, 17, 19, 23, 14, 15, 16, 2, 20, 7, 18, 8
sand
a: 22, 7, 6, 2, 3, 30, 31, 21, 0, 16, 19, 17, 23, 5, 12, 13, 14, 15, 10, 11, 8,
9, 4, 25, 28, 27, 24, 1, 20, 18, 29, 26
b: 18, 16, 1, 17, 9, 26, 24, 8, 0, 4, 25, 12, 19, 14, 6, 22, 20, 21, 5, 13, 15,
23, 7, 3, 2, 10, 11, 27, 31, 29, 28, 30
styr
a: 0, 8, 13, 24, 28, 29, 9, 5, 22, 4, 11, 19, 2, 18, 16, 26, 7, 12, 14, 20, 15,
3, 27, 17, 21, 31, 23, 25, 6, 1
b: 9, 25, 24, 10, 4, 28, 26, 30, 27, 19, 13, 29, 1, 0, 11, 12, 3, 18, 22, 6, 23,
2, 21, 16, 8, 31, 5, 20, 17, 14
ex1
a: 0, 4, 2, 16, 12, 18, 6, 15, 26, 11, 23, 19, 27, 30, 3, 31, 7, 5, 22, 1
b: 0, 1, 4, 8, 3, 5, 12, 11, 23, 7, 28, 22, 18, 15, 19, 14, 21, 6, 17, 2
ex4
a: 11, 3, 6, 15, 1, 4, 0, 9, 12, 14, 8, 10, 13, 5; b: 4, 2, 1, 3, 6, 13, 0, 5,
15, 7, 12, 10, 11, 9
opus
a: 6, 2, 11, 0, 15, 8, 10, 4, 5, 1; b: 9, 11, 15, 0, 7, 8, 12, 2, 1, 10
train11
a: 0, 1, 2 ,3, 7, 9, 8, 10, 14, 6, 4; b: 0, 1, 2, 3, 7, 9, 8, 10, 14, 6, 4
Chapter 4
Differential CMOS single

edge-triggered fli p-flop
Flip-flops are the basic building blocks of synchronous digital circuits and, to a large extent,
determine circuit power dissipation. Hence, low power design techniques of flip-flops are essential
to design low power sequential systems. As a consequence, many genius techniques have been
recently proposed to reduce the power dissipation of the flip-flops [49][19][68][102][100][54][72].
Categorized by the input sampling, edge-triggered flip-flops can be grouped into two types:
pulse sampling [102] and level sampling [100]. The former only needs one latch and hence has
a simpler structure, however, it suffers from long clock chain used to generate narrow pulses.
The latter needs two latches or flip-flops and hence needs more transistor count, which results
in the increase of its area and power dissipation [102]. Structurally, they can be categorized
into differential and non-differential flip-flops. The structure characteristic of differential flipflops is that they are composed of one differential amplifier, flip-flop cell, and some auxiliary
circuits. Hence, differential flip-flop has some advantages over the non-differential one since it
has complementary outputs, can amplify a small voltage signal and at the same time latch data.
Much effort has been paid to improve its performance [54][59][72]. In [54], reduced clock-swing
flip-flop is proposed and power saving is achieved by lowering the voltage swing of the clock
system. However, a backgate bias is needed, which complicates the power supply. In [100], a
57
CHAPTER 4. DIFFERENTIAL CMOS SINGLE EDGE-TRIGGERED FLIP-FLOP
58
flip-flop design with clock-gating on both master and slave latches is presented. However, the
area penalty for the additional circuit is very high and the power saving is significant only when
the switching activity of the input signal is very low. Looking at the published circuits, their
designs suffer from two disadvantages.
1. For the version clocked with NMOS and PMOS transistors in flip-flop cell, clock load
is heavy and static power dissipation is significant if reduced-swing clock singles are applied to
PMOS transistors [54][59].
2. For the version clocked with NMOS transistors in flip-flop cell, it suffers from long clock
chain, which consumes significant clock power [72] .
4.1
Differential CMOS single edge-triggered flip-flop
In [59], two N - C 2 MOS output latches were used to improve the speed of flip-flops but this
is traded with power dissipation penalty. There are two disadvantages for this design. One is
that the clock load is too heavy because it uses many clocking transistors. The other is difficult
to apply for reducing the signal swing due to clocking with NMOS and PMOS transistors. To
overcome these problems, in [72], a differential CMOS Single Edge-Triggered Flip-Flop (SET-FF)
based on clock racing was proposed.
The flip-flop in Fig. 4.1 consists of a differential amplifier called a cell and a clock chain
which is composed of three inverters [72]. The clock chain generates a delayed CPl. CP and
CPI are applied to the cascaded transistors, m7 and m8, respectively, which generate a narrow
pulse at the rising edge of each clock pulse. Hence, the flip-flop is triggered at the rising edge of
the clock pulse. Transistor sizes, which are expressed in W / L, are marked beside the inverters.
To obtain proper width of the narrow pulse, transistor sizes, Wand L, of the second inverter in
the clock chain are the double of the other inverters. If D, Q and Q' are used to represent the
input, present state and next state signals respectively, the next state equation can be expressed
into
= DCicp + QCicp
59
(4.1)
Here, Cicp represents the rising edge of clock pulse.
Q
Q
D
.1
~P1
~1
-::-
4/1
CP1
CP
(a)
DQQ
CP
(b)
Figure 4.1: Differential CMOS SET-FF (a) Circuit implementation; (b) Logic symbol
Because the cell is driven by a narrow pulse, the output latches can be removed and only two
NMOS transistors are clocked. Hence, this design has a simpler structure and a feature to apply
for data and clock signals with reducing swing. The experimental results show that it has better
performance in power and speed than those used in commercial processors. On the other hand,
the advantage of edge-triggered flip-flops is that the setup time for data input is independent of
the clock pulse width. It is also less sensitive to noises. This makes system design simpler.
4.2
Circuit parameter optimization
For a flip-flop, speed and power dissipation are two key parameters. However, they mainly
depend on the circuit structure and technology parameters. For a specific circuit structure,
power delay product (PDP) is used to measure its performance. Therefore, circuit optimization
60
aims to optimize PDP.

Definition 4.1. Power Delay Product (PDP) is defined by the following equation:
PDP = T delay
* Pw
(4.2)
Here, Tdelay is the delay of a flip-flop while Pw is its power dissipation.
4.2.1
Delay
For the clarification, it is helpful to give some definitions here.

Definition 4.2. DCQ : Clock-to-Q time, the propagation delay from the C to the Q, assuming
that the D signal has been set early enough relative to the triggering edge of the clock pulse.
Here C refers to the Clock.
Traditionally, the delay of flip-flops is referred to Clock-to-Q delay. However, some authors
pointed out that using DCQ delay as a relevant performance parameter for a flip-flop is misleading
because DCQ delay does not take the setup time into account [1][99].
Definition 4.3. Ts : Setup time, the minimum time between a D change and the triggering
edge of the clock pulse such that the output Q will be guaranteed to change so as to become
equal to the new D value.
The setup time is the delay between the data input of the flip-flop and the storage element
as the data takes a finite time to travel to the storage point.
Thereby, it is proposed to use D-to-Q as the delay of a flip-flop. This is defined as follows:
TVQ
= DCQ +Ts
In this thesis, we will accept both DCQ and TVQ to measure the delay of a circuit.
For the measurement of TVQ, the methodology proposed in [1] is followed.
(4.3)
CHAPTER 4. DIFFERENTIAL CMOS SINGLE EDGE- TRIGGERED FLIP-FLOP
4.2.2
61
Circuit optimization
For a high-performance and low-power application, speed and power are equally important.
Usually, there is tradeoff between power and speed. To compromise the two, the concept of the
minimum power-delay product (minPDP) is used. The minimum power-delay product is the
optimal energy utilization at a given clock frequency.
Take Fig. 4.1 for example to explain this. Fig. 4.2 is the same version of Fig. 4.1. The main
difference is that all inverters are substituted by nMOS and pMOS transistors. However, it is
not easy to optimize PDP because theoretically PDP is a multiple variable function of circuit
parameters. Given the variety of designs, it is not always simple to express the PDP as a function
of one common variable. For simplification, we take the transistor channel width as a common
variable. Then, delay, power and PDP can be expressed into functions of one variable. Set the
minimum transistor length to 1/1- and channel width is marked beside the transistors using w.
Some transistors are pre-optimized based on the parameters referred to in reference [72].
Q}
I~Q
-=-
1~D
-=-
-=-
CP
~~
/2 ~~/2
wl2
-=-
CPl
w/2
-=-
Figure 4.2: Power delay product optimization

Fig. 4.3 shows the linear relationship between the power and transistor width w while Fig.
4.4 indicates the nearly inversely proportional relationship of delay versus width w. Fig. 4.5
shows the curve of PDP versus width w, which shows that there is a point of the optimum
62
power-delay tradeoff. Based on this analysis, the circuit parameters can be optimized to the
point of minimal PDP.
:s:-:J
-s:::
0
:J:J
~
.9U1I
U)
.-
"
10-
G.I
i:
1400
1200
1000
800
600j
400
.-A
200
0
f1.
0
0
10
20
30
40
Chanrlel width (urn)
Figure 4.3: Power versus width w

The procedure presented above is used to optimize circuits. Although the procedure could
be developed into an automatic tool, it is not our main point here. Further details can be found
in reference [99].
4.3
Differential CMOS Single Edge-Triggered Flip-Flop with ClockGating (CG-SETFF)
As known, clock signal is the most frequent transition signal in the circuit. In one clock cycle,
clock signal has two transitions while other signals in the circuit (exclude glitches) have one
transition at most. Hence, in the clock chain, transition is the most frequent and correspondingly
it consumes significant power. According to the working principle of D flip-flop, next state Q'
follows the input signal D. If D does not change, the flip-flop does not need to be triggered. If we
63
can detect the idle conditions of input signal D and stop the clock signal, then power dissipation
can be saved. Based on this idea, a new design is proposed in this section.
In Fig. 4.1, it is observed that whether the input signal D varies or not, there is always
a narrow pulse to trigger the flip-flop. However, when the input signal D is in low switching
activity, it is unnecessary to trigger the flip-flop for each clock pulse. For example, when D keeps
at 1, the flip-flop does not need to be triggered. In this case, those narrow pulses to the flip-flop
are redundant and the transitions that happen in the clock chain are also redundant. If the clock
of the clock chain can be deactivated, power can be saved. Hence, if a non-redundant transition
clock chain can be designed, then it can be expected to save significant power when the D input
transitions are low.
Fig. 4.6 is a schematic diagram of a non-redundant transition clock chain. In Fig. 4.6, D and
Q are the input and the output signals of a flip-flop respectively and CP is its clock. XOR gate
is used to compare the input D and the output Q. When D=Q, the output of the XOR is 0 (1
when D#Q). The output of the XOR gate is used as an enable signal for the tri-state inverter.
-.
1.5
U1I
s::::
i;'
0.5
0
0
. .
10
20
30
40
Channel width (um)
Figure 4.4: Delay versus width w
64
The clock chain is only driven by OP when D#Q. Hence, compared to the conventional flip-flop,
redundant OPs are eliminated. If CP1 and CP2 are used to control the gates of m7 and m8
the narrow pulses generated are also non-redundant. Since the narrow pulse is generated at the
falling edge of OP, the flip-flop will sample the input signal at the falling edge of OP. Using the
dual trail signals D and Q from the flip-flop, one possible circuit implementation is shown in Fig.
4.7 [125]. mg-m13 make up a control circuit, which is used to selectively pass OP to the clock
chain. When D
= Q,
mll is off and OP is blocked to the clock chain. When D # Q, inverted
OP is passed to the clock chain.
0.6 ...-,- - - - - - - - - - - - - - - ,
0.5
~
0.4
.-
0.3
"5
u
:s 0.2
~ 0.1
a..
O+I------~----~------~----~
10
20
30
Channel width (urn)
Figure 4.5: PDP versus width w
CP2
CP
CPl
Figure 4.6: Non-redundant transition clock chain
40
4.4
65
Circuit simulations and power dissipation measurement
Circuit simulations are conducted using PSpice II with Level 3 at 5V, IlL technology and 50MHz
clock. SPICE parameters for IlL technology are shown in Table 4.1.
Table 4.1: SPICE parameters of MOS transistors for a generic 1.0ILm process
.MODEL CMOSN NMOS LEVEL=3 PHI=O.6 TOX=2.03E-08 XJ=O.15U TPG=1
+VTO=O.7333 DELTA=9.445E-Ol LD=1.0E-09 KP=1.2964E-04 UO=762.1 THETA=5.246 E-02
+RSH=2.3650 GAMMA=O.4481 NSUB=1.75E+16 NFS=2.356E+12 VMAX=1.487E+05
+ETA=1.485E-Ol CJ=1.1962E-04 MJ=0.4398 CJSW=4.6953E-I0 KAPPA=9.51E-02
+CGDO=2.5516E-12 CGSO=2.5516E-12 CGBO=3.0108E-I0 MJSW=O.123994 PB=O.8
.MODEL CMOSP PMOS LEVEL=3 PHI=O.6 TOX=2.03E-08 XJ=O.15U TPG=-1
+VTO=-O.9679 DELTA=4.3070E-Ol LD=1.0E-09 KP=4.3207E-05 UO=254 THETA=1.7060E-Ol
+RSH=2.5530 GAMMA=0.497 NSUB=2.153E+16 NFS=4.566E+12 VMAX=1.82E+05
+ETA=1.8290E-Ol KAPPA=3.225 CGDO=2.5516E-12 CGSO=2.5516E-12 CGBO=3.5207E-I0 i
+CJ=5.3093E-04 MJ=O.5074 CJSW=7.8757E-ll MJSW=O.077193 PB=O.85
I
The transistor lengths are minimum for the above technology, while the transistor widths
Q
Q
D
"
_I
CPl
C~
CPICP2
L
CP2
D8
CP
(a)
Q
Q
(b)
Figure 4.7: Differential single edge-triggered CMOS flip-flop with clock-gating (CG-SETFF)(a)
Circuit implementation; (b) Logic symbol
66
are Sf.t except m3, m4, m7, mS, m12, m13 and those in the clock chain. The widths of m7 and
mS are l2f.t. The sizes of m3 and m4 are: W
are: W
4f.t, L
2f.t, l
4f.t while the sizes of m12 and m13
If.t. The transistor sizes of the clock chain, which are expressed in W / L,
are indicated beside the inverters in Fig. 4.1. These parameters are determined for optimizing
both speed and power as stated in Section 4.2.2. The PSpice simulation shows that the proposed
flip-flop has the correct behavior as shown in Fig. 4.S.
5 .GU
T --1
CP
GU
5 . GU
4-1---.1
T - - - - - - ,---, -
.,..---..,. -
_.-
-,,- - - -.,---------------,-
GU +------'5 .GU T---------ry---,-------v-
---
---
-,--'--
--------
---------
G -+I---------l
Gs
---
-I- - - -
L -_ _ _ _ _ _L
_ _ _ _ _ _ _ -, _ _ _ _
Time 5Gns
tGGns
Figure 4.S: Transition behavior of proposed flip-flop

Average power dissipation is measured when each output is loaded with CL
[72]. The power consumption is calculated as:
= O.lpicofarad
Pw
=; 1Tm{
67
(4.4)
VDDIDD
Here Tm is the measured time period of the D signal.

The ideal input signal D and clock signal CP are buffered by two series inverters. The test
bench for measuring power dissipation is shown in Fig. 4.9. The reason why this test bench is
chosen is to provide realistic data and clock signals, which themselves are fed from ideal voltage
sources. Capacitive load at the data input simulates the fanout signal degradation from the
previous stages while capacitive load at the outputs simulate the fanout signal degradation caused
by the succeeding stages. The circuit with heavy load is to estimate its driving capabilities. To
estimate the power consumption, the simulation is performed by varying the switching activity
of the D input. The switching activity is defined here as the average number of transitions of
D in a clock cycle. The maximum switching activity for a glitch free signal is 1 while switching
activity could be greater than 1 if the presence of glitches is taken into account. The switching
activity for a signal that behaves as clock signal is equal to 2.
Power consumption is reported as a function of input signal switching activity, assuming
that the input signal has no glitches. Six input cases are considered for the switching activity
of 0, 0.2, 0.4, 0.6, 0.8, 1.0. The methodology proposed in [99] is followed to measure the power
dissipation. The power spent on the capacitative load needs to be excluded to get a fair picture
of the circuit's power behavior. For the given load CL
O.lpicofarads, that portion of power
reaches the values presented in Table 4.2.

D
ICL
CP
ICL
Figure 4.9: Test bench for measuring the power dissipation of flip-flops
The power dissipation is evaluated for Tm
= 400ns. The measure statement of average power
68
dissipation in PSpice is used to measure the interested power dissipation.

For a fair comparison of the proposed circuit and that reported in [72], the same PSpice
simulator and 1f.,l technology are used to simulate the circuits. The result of simulations is shown
in Fig. 4.10. From Fig. 4.10, it can be seen that the proposed flip-flop consumes lower power
than the conventional one when the D input switching activity is < 0.65. When D is idle, the
power dissipation of the proposed circuit is only 14% of that of a conventional flip-flop. The
reason is that when D is idle both the cell of the flip-flop and the clock chain are shutdown and
hence the dynamic power dissipation is nearly zero. However, when the switching activity of the
input signal is greater than 0.65, the extra devices consume the significant power. As a result,
the total power is greater than that of the conventional one .
..-.. 500
:500
.9 400
~
.9300
i'/)
200
100
g,
0.4
0.2
D inplJt
SI).I
0.5
0.8
itohing activity
Figure 4.10: Power dissipation against D input switching activity. P1: Proposed flip-flop; P2:
Conventional flip-flop
Table 4.2: Power dissipation for load capacitors versus switching activities
p.out -- a /coutVdd
2
f = 50MHz, Vdd = 5V
a = 1.0
a=0.8
a=0.6
a = 0.4
a=0.2
a=O.O
Differential structures
Cout = 2CL = 0.2picofarad
125f.,lW
100f.,lW
75f.,lW
50f.,lW
25f.,lW
Of.,lW
4.5
69
PDP measurement
To obtain PDP defined in equation 4.2, the delay needs to be measured. The methodology in
[1] is followed and the test bench in Fig. 4.11(a) is employed. The principle behind this method
is that the test bench forms a ring oscillator. To measure the time figures for two flip-flops,
the frequency of the clock pulse is gradually increased up to a point where the oscillator fails to
operate. This is the minimum clock period of the flip-flop, which has:
Tmin
= DCQ +Ts +DL
(4.5)
lll-.t 1-- -.t 1-- Dr.2
Q
Q
CP
Cf>
(a)
CJU - I
(b)
Figure 4.11: Timing test circuit (a) Bench circuit; (b) Timing diagram
Here, DCQ and Ts are as defined previously, DL is the delay of the combinational circuit.
The delay of the flip-flop can be obtained as long as DL can be determined. From the test
bench, it is known that DL is the inverter chain delay. Fig.4.11(b) shows the timing diagram.
Then, to find the delay of the flip-flop, DLl and DL2 need to be subtracted from the clock
period. We measured that the delay of the conventional flip-flop is 0.355ns while the delay of the
proposed circuit is 0.550ns. Then, PDP can be calculated based on the definition in equation
4.2. Fig. 4.12 presents the relationship of PDP versus the switching activities for the proposed
flip-flop and the conventional flip-flop. It can be seen that if using PDP to compare, the proposed
flip-flop has higher energy efficiency than the conventional one when the switching activity of
the input signal is lower than 0.4.
4.6
4.6.1
70
Low power binary counters

The proposed structures are suitable for applications with reduced switching activity. A typical
application is in synchronous counter design. In a synchronous counter, the input switching
activity in each flip-flop is different from place to place and is known beforehand. Hence, if the
proposed flip-flop is used in the low switching activity bits of the counter, a low power counter
can be obtained. Take a 4-bit binary counter as an example. It shows as follows.
If A, B, C and D are named for four flip-flops from the Least Significant Bit (LSB) to the
Most Significant Bit (MSB), then the state table for this counter is shown in Table 4.3. D, C,
B and A are used to represent present state variables while D', C', B' and A' to represent next
state variables.
0.4
0.3
0.2
Q)
.~
-t
0.8
'1
A
~ 01
0
0
0.2
0.4
0.6
Switching activitv
Figure 4.12: PDP versus switching activities
71
Assuming that D flip-flops are used, the following equations are obtained.
DA=A
DB=AEBB
Dc
(4.6)
= CEBAB
DD =DEBABC
Hence, using D type SET-FFs, the counter can be implemented in Fig.4.13. From the working
behavior of the counter, it can be known that the switching activity for each flip-flop input is
2- k for the kth bit, that is,
aA =
1.0,
aB =
0.5, ac = 0.25 and aD = 0.125. It can be seen
that except the lowest-order bit ( with switching activity 1.0) the other bit switching activities
are less than and equal to 0.5. In Section 4.4, it shows that using the proposed flip-flop can save
power if the input switching activity is less than 0.65. If we use the SET-FF for the lowest bit
while the proposed flip-flops (CG-SETFFs) are used for the remaining bits, then a lower power
counter can be obtained. However, for convenience, we use four CG-SETFFs instead.
Table 4.3: State table for a

Present state
Count D C B A
0
0 0 0 0
1
0 0 0 1
2
0 0 1 0
3
0 0 1 1
4
0 1 0 0
5
0 1 0 1
6
0 1 1 0
7
0 1 1 1
8
1 0 0 0
9
1 0 0 1
10
1 0 1 0
11
1 0 1 1
12
1 1 0 0
13
1 1 0 1
14
1 1 1 0
15
1 1 1 1

Next state
D
A
C B
1
0
0
0
0
1
0
0
1
1
0
0
0
1
0
0
1
1
0
0
1
0
1
0
1
1
0
1
1
0
0
0
1
1
0
0
1
1
0
0
1
1
1
0
1
1
0
0
1
1
1
0
1
1
1
0
1
1
1
1
0
0
0
oI
72
In order to measure the power dissipation, NAND gate and XNOR gate need to be implemented as well. The traditional design for 2 input and 3 input NAND is employed while for
XNOR, the low power design in [118] is used, which are shown in Fig. 4.14 for convenience. All
transistor sizes are W = 8/-lm, L = l/-lm.
Once again, the above two circuits are simulated using PSpice with level 3 at 5V, 1/-l technology and 50MHz clock. The power dissipation was evaluated for Tm = 400ns. The experiment
shows that 24% of the power is saved using the proposed flip-flop to implement the counter.
However, the saving varies with the number of bits. Because the circuit has a complicated combinational section, it is expected that the saving significance will be reduced with increasing the
bit number of the counter.
4.6.2
Binary twisted ring counters
It is well known that binary ring counter has simple combinational circuit while there are many
redundant transitions. If the proposed flip-flop is used in these applications, the power saving will
be much more significant than binary counters because the latter has complicated combinational
circuits. Using SET-FFs and CG-SETFFs, we implemented 4-bit, 6-bit and 8-bit twist ring
counters. Then, the same parameters as in Section 4.6.1 are used to simulate the circuits. Fig.
4.15 shows the power comparison results for the SET-FF based and CG-SETFF based twist ring
CP
Figure 4.13: SET-FF based binary counter
73
counters. It can be seen that the counter, using the proposed flip-flops to build, has much smaller
power than that using the conventional flip-flops. This is because there are many redundant
transitions in the twisted ring counter, which are blocked by new flip-flops. Furthermore, it can
be expected that the power savings become even more significant with the increase of the counter
V DD
out
out
c
b
D(a)
(b)
Figure 4.14: Gate circuit implementation (a) 2 input NAND and its logic symbol; (b) 3 input
NAND and its logic symbol; (c) XNOR and its logic symbol
74
bit number. The reason is that the redundant transitions increase with the bit number of the
counter. For example, a possible state diagram for a 4-bit twisted ring counter could be as in
Fig. 4.16, which has eight states. For each state transition between two states, only one bit
needs to be changed while the other three stay at the same state. The redundant transition rate
is 75%. For an 8-bit twist ring counter, it could be up to 87.5%.
~ 2500
..-
------+
-; 2000
Q
-:; 1500
c..
:~ 1000
....Q) 500
3:Q
""C
C.
4-b its
6-b its
P2
8-b its
bits
Figure 4.15: Power dissipation for twisted ring counters: P1- SET-FF based counters; P2-CGSETFF based counters
Figure 4.16: State diagram for a twisted ring counter
4.7
75
Summary
A non-redundant transition clock chain for differential CMOS single edge-triggered flip-flops
is presented. The experimental results show that compared to the conventional flip-flop, the
proposed flip-flop can achieve significant power savings when the D switching activity is < 0.65.
Reduction of power consumption can be as high as 86% when the input is idle. However, using
PDP to measure them, the proposed flip-flop has higher energy efficiency than the conventional
one when the D switching activity is < 0.4. Since the dual trail signals are provided by differential
edge-triggered flip-flops, the additional circuit to implement non-redundant transition chain is
simpler. In this design, the increased transistor count is 25% while in [100] it is 144%. Hence,
the clock-gating technique is more suitable to differential edge-triggered flip-flops with clockracing than to level sampling flip-flops. The comparison is also made for some applications. The
result shows that significant power savings can be obtained. However, the saving varies from
application to application.
Chapter 5
Differential CMOS double

edge-triggered flip-flops
In a digital system, the sequential part is the main contributor to power dissipation. The reason
is that one of the inputs to a sequential circuit is the clock pulse (CP) signal. Each CP has one
rising transition and one falling transition in one CP period. Hence, the switching activity is 2.
However, other signals in the circuit have at most one transition in one CP period if the glitches
from signal racing are not taken into account. On the other hand, the load of CP is alway the
highest in a digital system. For example, to distribute CPs and control the skew of CPs, one
needs to build a clock tree which consists of buffers, which results in increasing the total node
capacitance of a CP tree. As a result, research confirms that in a digital system 20-45 % of the
power is consumed by CPs.
However, traditional flip-flops are only triggered by a specific edge (rising edge or falling
edge) of a CP, which is called Single Edge-Triggered Flip-Flops (SET-FFs). On the other edge,
it is just wasteful to charge or discharge the capacitive load of the global clock line in a system.
This is particularly true in CMOS because the dynamic power is the dominant power. Therefore,
reducing the CP power consumption should reduce the total power dissipation for a system. If
flip-flops can be triggered by two edges, then half of power dissipation from CPs can be saved.
This leads to the design of double edge-triggered flip-flops [49].
76
CHAPTER 5. DIFFERENTIAL CMOS DOUBLE EDGE-TRIGGERED FLIP-FLOPS
5.1
77
Previous work
The double edge-triggered (DET) idea was presented as early as 1981 [110]. At that time,
it targeted high performance for flip-flops. With the increasing size of VLSI systems, power
dissipation gained the equal priority to area and speed. The double edge-triggered idea was
developed to reduce the power dissipation of VLSI systems [49][19].
Based on input sampling mode, DET flip-flops can be classified as level sampling and pulse
sampling. Level sampling DET flip-flops can be implemented using two latches[49][19]. Two
latches sample and store input signal alternatively under different clock phases. Pulse sampling
flip-flops usually consist of one latch and one clock chain [102]. The clock chain is used to
generate narrow pulses. The input signal is sampled during each narrow pulse.
For differential DET-FFs, level sampling version has been proposed in [68][58]. The disadvantage of DET flip-flops has been the substantial increase in the number of transistors required
to build such flip-flops compared to single edge-triggered (SET) flip-flops. For example, a typical SET-FF needs 16 transistors while DET flip-flop proposed in [58] needs 28 transistors. A
recently proposed circuit in [68] needs 26 transistors, which is shown in Fig.5.1 and is called
BALCS-DETFF.
.,
CK
----J>o---- CK
Figure 5.1: Differential CMOS DET flip-flop proposed in [68]: BALCS-DETFF
5.2
78
Clock Chain Based Double Edge-Triggered Flip-Flop (CCBDETFF) using single latch
Single-Edge-Triggered Flip-Flops (SET-FFs) change their states either at the rising edge or at
the falling edge of the clock pulse. Take the differential CMOS SET-FF proposed in [72] for
example. For convenience, it is redrawn in Fig. 5.2, which consists of a flip-flop cell and a clock
chain. The clock chain generates a delayed CPl. Then, CP and C PI are applied to the cascaded
transistors, m7 and m8, respectively, which generate a narrow pulse at every rising edge of each
clock pulse. Hence, the flip-flop is triggered at the rising edge of the clock pulse. However, this
design suffers from one disadvantage. The flip-flop responds only once per each clock pulse cycle.
The other of the two clock transitions accomplishes nothing for the flip-flop. But this transition
may cause transitions of some logic gates within the flip-flop, which consumes power.
Q
Q
"
LrcP1
4/1
-=-
~1
812
CP1
CP
Figure 5.2: Differential CMOS single-edge triggered flip-flop (SET- FF)

To solve this problem, there are two solutions.
transitions, which is proposed in [125].
One is to refrain from redundant clock
The other is to use the other transition to change
states. With both edges able to cause state transitions, the redundant clock transitions can
be eliminated. Moreover, the clock period can be shortened since there is no need to wait for
the clock signal to toggle. This is the design principle of Double Edge-Triggered (DET) scheme
which enables a halving of the clock frequency and hence reduces power dissipation [49].
5.2.1
79
CCB-DET flip-flop
From Fig. 5.2, a clock chain consisting of three inverters can generate a delayed CP, CP3 . If
we use four inverters to implement a clock chain, then four delayed CPs can be generated. The
timing diagram is shown in Fig. 5.3 (b) where CP4 is a delayed CPl. If CP and CP3 are logically
ANDed while C PI and C P4 are ANDed, then two narrow pulses corresponding to two edges of
a clock pulse can be obtained as shown in Fig. 5.3. To even two narrow pulse width, W f L of
transistors in the clock chain is taken as indicated beside four inverters. This can be proved as
follows.
If
tl
is the delay of a 4fl-inverter while
t2
is the delay of a 8f2-inverter. Then the delay
between CP3 and CP is:
tcpa-cp
= tl + 2t2
Also, we have the delay between CPI and CP4 :
tCP4-CPl
= 2t2 + tl
Hence, two narrow pulses have the same pulse width.

The proposed double edge-triggered flip-flop cell is as in Fig. 5.4. There are two branches,
which are composed of four transistors, m7, m8, m9 and mlO, at the common source terminal
of the amplifier. CP and CP3 are used to control the gates of m7 and m8 to generate a narrow
pulse corresponding to the rising edge of clock pulse while CPI and CP4 are used to control the
gates of m9 and mlO to generate the other narrow pulse corresponding to the falling edge of
clock pulse. In this way, the flip-flop can change its state triggered by two clock transitions.
5.2.2
80
Circuit simulations
Circuit simulations are conducted using PSpice II with Level 3 at 5V, 1.0J-tm process. The
transistor lengths are minimum for the above technology, while the widths are 8J-tm except those
specified as follows. The widths of m7, m8, mg and mlO are l2J-tm while the sizes of m3 and
m4 are W
2J-tm, L
4J-tm. The transistor sizes of the clock chain, which are expressed in
W / L, are indicated beside the inverters in Figs. 5.2 & 5.3. These parameters are determined
cJL
c~
c~
c~
c~
411
8/2
812
cp
c~
(a)
(b)
Figure 5.3: Double edge pulse generation (a) Clock chain; (b) Timing diagram
Q
Q
D
.1
D-1l-Q
cP~Q
-::-
(a)
(b)
Figure 5.4: Clock Chain Based double Edge-Triggered Flip-Flop (CCB-DETFF) cell (a) Flip-flop
cell; (b) Logic symbol
81
for minimum power as stated in Section 4.2.2. The PSpice simulation shows that the proposed
flip-flop has correct transition behavior as shown in Fig. 5.5. In Fig. 5.5, the transient behavior
of the single edge-triggered flip-flop is also shown for comparison.
The test bench in Fig. 4.9 is used to simulate circuits. Circuits in Figs. 4.7 & 5.2 are used to
be reference circuits. Once again, six input cases, the switching activity of 0, 0.2, 0.4, 0.6, 0.8,
1.0, are considered to test the power dissipation of circuits. However, to get the same output
throughput as the double edge-triggered flip-flop does, the clock frequency to the SET flip-flops
should be 100MHz while the clock frequency for the DET counterpart is 50MHz. The results
are shown in Fig. 5.6 and Table 5.1.
Table 5.1 shows the power saving rate compared to SET-FF in Fig. 5.2. From Fig. 5.6, it
can be seen that CCB-DETFF always has lower power dissipation than SET-FF with different
input switching activities. However, the saving rate varies with the input switching activity
From Table 5.1, when
0:
0:.
= 0, the saving rate is up to 44.7%. Ideally, the saving rate should
be 50% because the clock frequency is reduced half and the power is consumed mainly by the
clock chain. The reason which is not equal to 50% is that two cascaded transistors, m9 and m10,
and the extra inverter in the clock chain consumes the power. When
is reduced to 20.8% from 44.7% when
0:
0:
1.0, the saving rate
O. This is because the flip-flop cell also consumes
significant power and the weight of the clock power in the total power is decreased.
Table 5.1: Power savings compared to the circuit in Figs.4.7 & 5.2
0.2
0.4
0.6
0.8
1.0
Switching activities (0:)
0
0
SETFF (%)
0
0
0
0
0
86.5 54.2 28.4 5.5 -14.8 -30.1
CG-SETFF (%)
CCB-DETFF (%)
44.7 37.7 30.8 27.3 23.7 20.8
From Fig. 5.6, among three flip-flops, CG-SETFF has the lowest power dissipation when
0:
< 0.38 while CCB-DETFF has the lowest power dissipation when
0:
> 0.38. In addition,
SET-FF uses 20 transistors while CG-SETFF uses 25 transistors. However, CCB-DETFF uses
24 transistors.
CHAPTER 5. DIFFERENTIAL CMOS DOUBLE EDGE- TRIGGERED FLIP-FLOPS
S .1V
82
r-
ep
-a.1V
S.lV
o
-a.1V
S.lV
o
-a.1V
i- m'run---W-T
J----
mTu-r 1-
_,m___
----
mm
1
I
[Um[uJ -u---m--]_u_mJU] ___ r-]

as
Sans
laans
Til'lle
(a)
S.lV
CP
-a.1V
S.lV
D
-a.1V
S.lV
Q
-a.1V
~---[----~--I----[--~]----~---[--- [ ___ 1---]
m__ _\umu:
J_~]mnmITmLm-u] L__
j
[m-LJ ---m-----:lm ___ JuJ _ur-j
as
Sans
laans
Til'lle
(b)
Figure 5.5: Transient behavior (a) Differential CMOS SET-FF in Fig. 5.2; (b) Differential
CCB-DETFF in Fig. 5.4
5.3
83
Pass Transistor Based Differential CMOS Double Edge-Triggered

Flip-Flop (PTB-DETFF)
5.3.1
Voltage scaling property for CCB flip-flops
A common disadvantage for the above designs is that there are cascaded transistors in the
common source terminal of the differential amplifier (briefly called clocked transistor), which
introduces the extra resistance. This degrades the input threshold of the amplifier. This is
investigated by measuring the input threshold versus the gate voltage of the clocked transistors
in Fig. 5.2. The experiment is conducted as follows.
DC characteristic of the circuit in Fig. 5.2 is simulated using PSpice with 1 urn technology
and 5V power supply. The Input Threshold Voltage (lTV) is defined as the input voltage when
the output high level of the circuit is 2.4V. Varying the gate voltage of clocked transistors, the
lTV versus the gate voltage is measured. Then, reduce the number of clocked transistors in
Fig. 5.2 to one and repeat the above experiment. For the simplification, the former circuit is
i: 700
::::s
600
.-.....,0 500
~ 400
U1I
.-U1I 300
":10- 200
it:
QJI
!:
0
Q.
P3
100
0
0
0.2
0.4
0.6
0.8
S wit chi n 9 act iv it i e s
Figure 5.6: Power dissipation versus switching activities: Pl- result for Fig. 5.2; P2-result for
Fig. 4.7; P3-result for the proposed circuit
84
called Cir2ct while the latter Circt. The results are shown in Fig. 5.7. It can be seen that the
circuit, Circb which has one clocked transistor has lower lTV than the circuit, Cir2ct, which has
two clocked-transistors and furthermore the former is still workable when the gate voltage of the
clocked transistor is 2.0V .
___ Vlh- TV\O

3.8
-+- Vlh-Cne
~3.4
."
~ 2.6
:It
2.2
~ 1.8
!j1.4
e-
1.5
2.5
3.5
4.5
Gate voltage ct clocked

transistor (V)
Figure 5.7: Input threshold voltage versus gate voltage of clocked transistors. NB: Vth-Two
stands for threshold voltage of Cir2ct while Vth-One for that of Circt
This confirms that the cascaded transistors in the common emitter degrade the lTV of the
amplifier. From Fig. 5.7, it also shows that reducing the gate voltage of clocked transistors has
less effect for the lTV of Circt than for that of Cir2ct. The lTV will increase fast when the gate
voltage of the clocked transistors is reduced to 4.0V for Cir2ct. However, for Circt, the lTV
remains the same when the gate voltage is reduced to 2.0V. Hence, it is possible for Circt to
use low clock swing, which will significantly save the clock power. As known, dynamic power is
proportional to voltage squared. Hence, a good voltage scaling property is important to design
a low power differential flip-flop.
5.3.2
85
PTB-DETFF circuit
The proposed circuit is based on the following idea. If a narrow pulse corresponding to each
edge of clock pulse is available, the two cascaded transistors in the common source terminal in
Figs. 5.2 & 5.4 can be substituted by one transistor. Then the input threshold degrading can
be avoided. Furthermore, if two narrow pulses corresponding to two edges of the clock pulse are
generated to drive the transistor, double edge-triggered differential flip-flop can be implemented.
Narrow pulses can be generated by the clock-racing. A possible scheme is shown in Fig. 5.8.
There, C P1 is a delay C P and the delay time is determined by two cascaded inverters. If they
are input into an XOR gate, then narrow pulses (N Ps) corresponding to two edges of clock pulse
can be obtained. This is explained in Fig. 5.8.
CP
CP
CPl
NP
Figure 5.8: Narrow pulse generating scheme
Using transmission transistors, a possible circuit implementation is shown in Fig. 5.9 (a). All
branches are ANDed together as the output. From Fig. 5.8, CP & CP1 have four combinations,
namely 00,10,11 & 01, labelled as A, B, C & D respectively. As known, transmission of a logic
zero is not degraded through nMOS transistor while transmission of logic 1 is not degraded
through pMOS transistor[118]. Take Case A for example to analyze the working principle of the
circuit. For Case A, CP
=0
and CP1
= o. CP
is applied to the input of pass transistor mt4
while CP2 = 1, inverted CP1 , is applied to the gate of mt4. Then, logic zero is transmitted to
the output of mt4. The other cases can be analyzed similarly. However, it should be noticed
that narrow pulse width is widened because of the delay of inverter 3. Hence, when determining
the parameters of inverter 3, this factor should be considered.
Having this pulse generator, the flip-flop cell can be as shown in Fig. 5.9 (b) where only one
86
transistor m7 is required at the common source terminal. Because the generator only drives one
nMOS, the driving capability of Fig. 5.9 (a) is not a problem.
5.3.3
Circuit simulations
Circuit simulations are conducted using PSpice II with level 3 at 5V, 1{l technology and 50MHz
clock. The transistor lengths are minimum for the above technology, while the widths are 8{lm
except those specified as follows. The widths of m7 and m8 are 12{lm while ones of mtl, mt2,
mt3 and mt4 are 2{lm. The sizes of m3 and m4 are W
2{lm, L
= 4{lm. The transistor sizes
of the clock chain, which are expressed in W / L, are indicated beside the inverters in Fig. 5.9.
These parameters are determined for optimizing the power. The PSpice simulation shows that
the proposed flip-flop has correct transition behavior as shown in Fig. 5.10.
The test bench in Fig.
is followed.
4.9 is used to simulate the circuits and the same methodology
Average power dissipation is measured when each output is loaded with CL =
O.lpicofarad. Four input sequence cases are considered: Case 1, Case 2, Case 3 and Case 4.
Case 1 has an input change at every clock phase while Case 2 has an input change at every-other
clock phase. Case 3 and Case 4 are constant but with a glitch at every and every-other clock
Inv1
4/1
CP~
Inv 2
8/2
_~~D ~r!'l r-l>o-Q
Inv 3
411
)!~~~Q
,
_c
nDQ
CP
-=(a)
(b)
(c)
Figure 5.9: Proposed differential CMOS double edge-triggered flip-flop (DET-FF) (a) Narrow
pulse generator; (b) Double edge-triggered flip-flop cell; (c) Logic symbol
87
phase, respectively. The power dissipation is evaluated for a time period of 400 ns. The power
spent on output loads is excluded from the total power to get a fair picture of the power behavior
of the circuit.
The same PSpice simulator and 1tt technology are used to simulate SET-FF circuit given in
Fig. 5.2. However, to get the same output throughput as the double edge-triggered flip-flop, the
clock frequency to the single edge-triggered flip-flop should be 100M Hz.
Table 5.2 shows power reduction compared to the power dissipation of the circuit in Fig. 5.2,
which includes clock power dissipation. It can be seen that the proposed circuit has significant
power savings compared to the single edge-triggered counterpart. However, the power saving is
not equal to 50%. The reason is that three extra pass transistors consume power.
Table 5.2: Power dissipation comparison between SET-FF and PTB-DETFF
Input sequences
Case 1 Case 2 Case 3 Case 4
489.7
SET-FF in Fig. 5.2 (ttW)
240.2
383.3
279.2
PTB-DETFF in Fig. 5.9(ttW)
318.9
231.7
175.0
155.4
34.9
39.6
37.3
35.3
Power Reduction (%)
Table 5.3 shows the property of voltage scaling for two circuits. Case 2 is chosen as a test
sequence. It can be seen that the proposed circuit has better property of voltage scaling. The
S.tv
CP
-O.tv
5.1V
r---
~--~-- ?r--~~~~----~
I
I
I
I
-O.tV
-r..J
LJ
1...------1
OS.1V :-----~-----~----]- -O.tv
Os
- -SOns
-----1
tOOns
Time
Figure 5.10: Transition behavior of the propose circuit
88
proposed circuit can work well even if the supply voltage is as low as 1.2V while the circuit in
Fig. 5.2 does not work when the supply voltage is lower than 1.8V. It can also be seen that the
power reduction for the proposed circuit is up to 39.2% compared to the single edge-triggered
one under workable supply voltages. The area penalty is that PTB-DETFF needs 3 transistors
more than SET-FF.
Table 5.3: Voltage scaling property between SET-FF and
Supply voltage (V)
3.0
2.0
1.8
SET-FF in Fig.5.2(ttW)
120.7 50.3 40.6
PTB-DETFF in Fig. 5.9(ttW) 74.2 31.1 24.7
Power Reduction (%)
38.5 38.2 39.2
PTB-DETFF
1.5 1.2
16.6
6.9
Then, BALCS-DETFF in Fig. 5.1 is simulated using 1tt technology. The width of transistors
is 8ttm while the length is minimum. The experiment is carried out as follows. Under the specified
power supply voltage, the minimum clock swing is determined. Then the power is measured with
corresponding input switching activities. Two power supply voltages, 5.0v and 2.5v, are selected.
The results are shown in Table 5.4. In the table, the second row gives the minimum clock swing
for given power supply voltage. The third row gives the input switching activities while the
fourth row gives the measured flip-flops. The fifth row gives the power dissipation. The most
bottom row shows the power reduction compared to BALCS-DETFF.
Table 5.4: Voltage
scaling- property between BALCS-DETFF and PTB-DETFF
2.5
Power supply (V)
5.0
1.4
2.0
Mini Vclock swing (V)
Switching activities
a=0.5
a=0.5
a = 1.0
a = 1.0
DETFFs
BALCS I PTB BALCS I PTB BALCS I PTB BALCS I PTB
53.5 I 51.1
88.3 I 71.9
242.5 I 231.9
Power (ttW)
402.4 I 298
25.9
4.4
18.6
4.5
Power reduction (%)
Both flip-flops have the same minimum clock swing under the same power supply. With
= 1.0 and power supply 5.0v, the power reduction of PTB-DETFF can be up to 25.9%
compared to BALCS-DETFF in Fig. 5.1. The reason is that when a clock is "high" at the
voltage of Vclock, two clocked pMOS transistors in Fig. 5.1 do not switch off completely, leaving
leakage current flow through either of two clocked pMOSs. However, PTB-DET flip-flop has
no such problem. With a
= 0.5, the power reduction is 4%, which is much smaller than the
89
case a = 1.0. This is because PTE-DETFF has a long clock chain with pass transistor branches,
which consumes significant power and this is taken into account. When the input signal has lower
switching activity, the weight of the clock power cell in the total power is increased and dominates,
which results in the decrease of the power savings. This can be confirmed by experiments. Table
5.5 shows clock power and flip-flop cell power for BALCS-DETFF and PTB-DETFF. The rate
of clock power to cell power for PTB-DETFF is 0.69:1 when a
when a
= 0.5.
However, for BALCS-DETFF, it is 0.09:1 when a
1.0 while it goes up to 1.37:1
1.0 while it is 0.17:1 when
0.5. There are two solutions to further reduce the clock power. One is to refrain the
redundant clock transitions, which is proposed in [125]. The other is that a number of flip-flops
share a clock driver so that the clock power for each flip-flop is reduced [102]. Hence, if this
is taken into account, the power saving for PTB-DETFF is even more significant compared to
BALCS-DETFF.
Table 5.5: Power dissipation from clock chain and flip-flop cell with VDD = 5.0V and Vclock
5.0V
Switching activities
a = 1.0
a= 0.5
Power
Clock power Cell power Clock power Cell power
301.0
150.6
BALCS-DETFF (ttW)
26.4
26.2
192.4
97.9
131.9
133.8
PTB-DETFF (ttW)
Finally, it is noticed that PTB-DETFF only uses 23 transistors while BALCS-DETFF needs
26 transistors. Compared with BALCS-DETFF, the area saving of the proposed circuit is 11.5%.
5.4
Summary
Two differential CMOS double-edge triggered flip-flops, CCB-DETFF and PTB-DETFF, are
proposed in this chapter. CCB-DETFF could be the complementarity of the CG-SETFF to
design a low power system when the input switching activity is greater than 0.38. However,
further work needs to be done on how to design a system using both SET-FFs and DET-FFs.
PTB-DETFF is composed of a narrow pulse generator and a flip-flop cell. It only needs one
clocked NMOS transistor for the flip-flop cell. Hence, PTB-DETFF has good properties of
voltage scaling and clock swing scaling, which are very important to design a low power system.
Chapter 6
Multi-valued flip-flop approach

In the binary logic, the logic 1 is nearly set to the supply voltage Vdd. Each charging or discharging
to a node consumes ~vldCL' Hence, reducing Vdd is reducing logic level voltage and is an efficient
approach for low power design. However, there are some limitations to further scale down Vdd
as discussed in Section 1.2.2. If high supply voltage Vdd is still used and multiple logic levels
are set between 0 and Vdd, then logic level voltage between two adjacent logic values can be
reduced. Power dissipation for each charging and discharging to a node can be reduced. Take
quaternary logic for example. If Vdd = 5V, then four logic levels can be: 5V (VDD), 3.33V
(2VDD/3), 1.67V (VDD/3), and OV (Vss). Corresponding logic values will be {3, 2, 1, O}. Each
charging or discharging to a node transitioning between two adjacent logic levels only consumes
~(~Vdd)2CL' From the signal transmission and storage point of view, using multi-valued signal
can reduce transmission capacitance and result in power saving or improve transmission speed.
Fig. 6.1 shows a comparative result to transmit an eight bit binary signal 01110010 transmitted in
parallel and serial in binary and quaternary valued logic. Fig. 6.1(a) shows that using quaternary
signal only half of signal lines are needed and hence transmission capacitance can be reduced
while 6.1(b) shows that transmission time can be reduced to half.
Furthermore, multiple valued logic may resolve the serious pin-out problems encountered in
some VLSI circuit designs since if signals are used in four or more states rather than only two
the pin-outs could be reduced to less than half. Many logical and arithmetic functions have been
shown to be more efficiently implemented with multiple valued logic because fewer operations,
90
CHAPTER 6. MULTI-VALUED FLIP-FLOP APPROACH
91
gates, signal lines, etc. are required [31]. Yet, for multiple valued logic to be in wide use the
key problem is to design a less complicated multiple valued memory [112]. In recent years, many
ingenious multiple valued flip-flops (MVFFs) have been proposed [31][82][121][122][132].
Hence, multi-valued flip-flops may be a new approach to design low power flip-flops. However,
as a first step, the structure should be explored.
Looking at previous research, it can be seen that multiple valued flip-flop designs are analogous to binary flip-flops and have the following three characteristics.
(1) Structurally, MVFFs have similar characteristics to their binary counterparts and usually
are the extension of their binary version.
(2) Functionally, the next state behavior of the MVFFs are derived from their binary counterpart.
(3) In terms of working mode, the state transition is controlled by a binary clock pulse (CP),
which is similar to their binary counterpart as well.
The third characteristic contradicts the aim of using multiple valued signals to increase information content. Since all signals used in multiple valued circuits are multiple valued, the
clock signal should be multiple valued as well. For example, to a quaternary flip-flop, the clock
signal should be CP E {O, 1, 2, 3} rather than a binary signal. In order to increase information
content of the clock signal, ternary clock signal has been proposed by Wu [122]. There, besides
the standard binary clock signal, an additional signal is used to implement the preset function.
This effort, however, did not change the basic working mode of the MVFFs which are controlled
by a binary clock signal. In this paper, a new MVFF, controlled by a multiple valued clock pulse,
1---
0--1--1--1--0--0--1--0---
3 - - - CP
0--2 - - - xb
xq~
(a)
(b)
Figure 6.1: Comparison between binary and quaternary signal transmission. (a) Parallel transmission; (b) Serial transmission
92
is proposed.
6.1
Traditional Binary OP (BCP) controlled multiple valued master slave flip-flops in series
Similar to binary master slave flip-flops, the structure of multiple valued master slave flip-flops
is composed of two multiple valued latches in series and controlled by two inter-inverting binary
clock signals. The logic design of multiple valued latches can be classified by whether they have
transmission switches or not [133]. If not, the logic structure can be implemented by crossing
feedback from outputs to inputs. If yes, it can be implemented by employing the logic structure
of buffers with controlling switches. A quaternary CMOS master slave D flip-flop is proposed
by Prosser [82], which is composed of two quaternary buffers and two 2 to 1 multiplexers as
in Fig. 6.2(a). Here, D is the inciting input, Q is the output, CP is the clock pulse, and
Q, D E {O, 1, 2, 3}, CP E {O, 3}. Its working principal is: CP
= 3, the input of the master
latch receives the input signal D while the slave latch is in the storing state, and vice versa when
CP
= O. Hence, the state transition of the flip-flop happens at the falling edge of CP (3 -+ 0)
and the next state equation is Qf
= D. Because it is composed of two quaternary latches in series
and controlled by a binary CP, it is called a binary clock pulse (BCP) controlled quaternary
master slave flip-flop in series. Fig. 6.2 (b) shows a binary shift register, which is composed
of two D flip-flops of the type in Fig. 6.2 (a). The output of the shift register is expressed as:
Q"
= D. It means that after two falling edges of CP the output equals the original inciting input
signal D.
With this traditional design, a shift register of n digits is able to store n quaternary signals
in series, and is composed of 2n quaternary flip-flops controlled by a binary CPo
93
6.2
Multiple valued CP (MVCP) controlled multiple valued flipflops in parallel
Under the control of two levels in a CP, the master slave flip-flop shown in Fig. 6.2 (a) is in
receiving and storing states alternately so that its "invalid toggle" is avoided. This meets the
requirement of one state transition each CP for flip-flops. However, if two latches are connected
in parallel rather than in series, then one latch receives the input signal while the other stores and
outputs the stored signal. If CP is multiple valued such as a quaternary valued signal, the four
levels can be used to select four latches to work alternately, that is, one of them is in the receiving
state while the others are in storing states. This methodology is used to design multiple valued
CP
(a)
CP.
(b)
~- -[>2 to 1 multiplexer
Quaternary reshaper
(c)
Figure 6.2: BCP controlled quaternary master slave D flip-flop in series and two digit quaternary
shift register (a) BCP controlled quaternary master slave D flip-flops in series; (b) Two digit
quaternary shift register; (c) Logic symbols
94
CP (MVCP) controlled multiple valued parallel flip-flops which are proposed in this chapter.
Based on the above observation, a quaternary D flip-flop can be constructed as shown in Fig.
6.3 (a). It consists offour 2 to 1 multiplexers, four quaternary buffers, one quaternary threshold
function generator and one 4 to 1 multiplexer. The implementation of those components can be
found in the published literature [122][133][121]. Take the 4 to 1 multiplexer for example. If four
transmission gates are controlled by four outputs of a quaternary threshold function generator,
its circuit implementation is shown in Fig. 6.3 (b), where all inverters are binary.
">
10
4
to
2
3
CP
iCpi
to
2 1
3
CP
Quaternary 4 to 1 multiplexer
Threshold function generator
(a)
Figure 6.3: Logic structure of MVCP controlled multiple valued flip-flop in parallel (a) Logic
circuit; (b) Logic symbols
95
The working principal of the flip-flop in Fig. 6.3 (a) is: for any of the four values of the
quaternary C P, one of the four latches receives the input signal D while the others are at storing
states. The 4 to 1 multiplexer selects anyone latch at storing state and its stored state becomes
the output of the flip-flop. For the specific connection in Fig. 6.3 (a), when CP
receives the input signal and latch 2 outputs the stored signal; when CP
the input signal while latch 3 outputs the stored signal; when CP
can be output when CP
= 1, latch 2 receives
= 2, latch 3 receives the input
latch 1 outputs the stored signal. Hence, if the clock pulse changes from
0, latch 1
= 3, latch 4 receives the input signal and
signal and latch 4 outputs the stored signal; when CP
stored signal when CP
= 3. Let
-+ 1 -+ 2 -+ 3 -+ 0, the
Ql, Q2, Q3, Q4 be initial states
for four latches, Ql, Q2, Q3 and Q4 respectively. During the first three upward level transitions
of the CP, the outputs will be Q2, Q3, Q4, respectively. Therefore, the next state equation of
the D flip-flop shown in Fig. 6.3 (a) is expressed as:
Q"'
=D
(6.1)
The above equation states that the input signal will be output in three level transitions during
one CPo
6.3
Circuit implementation
The circuit implementation principle of all the components in Fig. 6.3 can be found in the
published literature [122][133][121]. However, their corresponding CMOS circuits will be implemented here based on transmission switch theory[122].
In the transmission switch theory, two algebraic systems are developed: Switching algebra to
describe switching state with variable E{T, F} and signal algebra to describe signal value with
variable E {O, 1,2, 3} if quaternary valued logic is taken for example. Two sets of connection
operations are introduced. The first set is for describing the physical process of how the signal
controls the state of an element. They are defined as follows.
96
Definition 6.1. High-threshold comparison operation
if x> t
'X: { :
(6.2)
if x < t
Definition 6.2. Low-threshold comparison operation
xt
== {
if x < t
F if x> t
(6.3)
where x is control signal and t E {a.5, 1.5, 2.5}.

The second set is for describing how the state of a MOS transistor decides whether the source
signal is transmitted to the drain or not.
Definition 6.3. Transmission operation
* a == {
if a = T
if a = F
<I>
(6.4)
where <I> describes the high-impedance state, c is called transmission source and a is the
switching state of a transmission switch network.
Definition 6.4. Union operation
Cj
H" #0, * C<2 :
Cl
* al if C2 * a2 = <I>
C2
* a2 if Cl * al =
<I>
(6.5)

In equation 6.5, the case CI
=I C2
97
and al = a2 = T is not allowed because a voltage conflict
arises between sources cland C2. In addition, the transmission operation
* takes priority over the
union operation #. Based on the above definitions and some derived properties, the function
expression can be obtained according to the logic functionality. Take a threshold 1.5 inverter
for example. The logic functionality is that if x < 1.5 the output will be 3 while if x > 1.5
the output will be O. Then, Taking threshold t
1.5, low-threshold comparison operation and
high-threshold comparison operation can be used to describe cases x < 1.5 & x
> 1.5. Hence,
based on definitions 6.4 & 6.5, the function expression is:
x(1.5) = 3 * x1.5#O *1.5 x
(6.6)
Take case x < 1.5 for example to explain the above equation. When x < 1.5, x1.5 = T while
1.5x
= F', since 1.5 x = F " O*1.5 x = <p' based on union operation,

x(l 5) =.
3*x1.5. since
x1.5 '
=T
,
x(1.5) = 3. According to the action principle of MOS transistors, when x < 1.5, a source voltage
of 3 must be transmitted by a pMOS transistor while when x
> 1.5 a source voltage of 0 must be
transmitted by an nMOS transistor. Hence, the circuit realization of the threshold 1.5 inverter
can be constructed as shown in Fig. 6.4. In Fig. 6.4, the number beside the transistor denotes its
on threshold. Thus, the circuit implementation with pMOS and nMOS transistors easily follows
based on the given function representation.
Similarly, the function expressions of all the logic components in Fig. 6.3 can be obtained as
follows:
Y2tol
Quaternary reshaper function
= Yo * x1.5#YI *1.5 X
(6.7)
x res
= 0 * xO. 5#1 *0.5 x1.5#2 *1.5 X2.5#3 *2.5 X
98
(6.8)
Threshold function Y x Y
YxY
={
0 Y =1= x
(6.9)
3 y= x
Here, x, y E {O, 1, 2, 3}. Then, the function expressions are:
0x o = 3 * x O.5#O
lXl
*0.5
= 3 *0.5 x1.5#O * (x O.5#1.5 x )
2X 2 = 3 *1.5 x 2.5#O
(x1.5#2.5 x )
3~_151-
x~ ~.
0-11.5
x(15)
I
(a)
-8>0--
x(1.5)
(b)
Figure 6.4: Threshold 1.5 inverter (a) Circuit implementation; (b) Logic symbol
(6.10)
(6.11)
(6.12)
3X3
= 3 *2.5 X#O * X 2.5
99
(6.13)
Y4tol
= Yo * XO. 5 #Yl
*0.5 X1.5#Y2 *1.5 X 2 .5 #Y3 *2.5 X
(6.14)
Their corresponding circuit implementations are shown in Fig. 6.5. However, for threshold
functions, only the implementation of xo is shown to save space. In addition, for quaternary 4
to 1 multiplexer, if four threshold function signals from four threshold function generators are
used, a simpler modified circuit can be obtained as shown in Fig. 6.5 (d). Then, the CMOS
circuit of the flip-flop in Fig. 6.3 can be constructed.
6.4
The proposed flip-flop in Fig. 6.3 has been simulated using PSpice with a supply voltage of 5V
for 0.5 f-lm technology. The SPICE parameters for a generic 0.5f-lm process are shown in Table
6.1.
Table 6.1: SPICE parameters for a generic 0.5f-lm process
.MODEL CMOSN NMOS LEVEL=3 PHI=O.7 TOX=10E-09 XJ=O.2U TPG=l
+VTO=O.65 DELTA=O.7 LD=5E-08 KP=2E-04 UO=550 THETA=O.27 RSH=2
+GAMMA=O.6 NSUB=1.4E+17 NFS=6E+ll VMAX=2E+05 ETA=3.7E-02
+CJ=5.6E-04 MJ=O.56 CJSW=5E-ll KAPPA=2.9E-02 CGDO=3.0E-10
+CGSO=3.0E-10 CGBO=4.0E-10 MJSW=O.52 PB=l
.MODEL CMOSP PMOS LEVEL=3 PHI=O.7 TOX=10E-09 XJ=O.2U TPG=-l
+ VTO=-O.92 DELTA=O.29 LD=3.5E-08 KP=4.9E-05 UO=135 THETA=O.18
+RSH=2 GAMMA=O.47 NSUB=8.5E+16 NFS=6.5E+ll VMAX=2E+05
+ETA=2.45E-02 KAPPA=7.96 CGDO=2.4E-10 CGSO=2.4E-10 CGBO=3.8E-10
+CJ=9.3E-04 MJ=O.47 C.JSW=2.9E-10 MJSW=O.505 PB=l
Four logic level voltages are set to: 5V (VDD), 3.33V (2VDD/3), 1.67V (VDD/3), and OV (Vss).
100
For all nMOS devices, W / L = 4p..m/0.5p..m, while for all pMOS devices, W / L = 8p..m/0.5p..m.
The PSpice simulation shows correct circuit operation as shown in Fig. 6.6.
0--------,
~ =fG>03::1-
T I ""/" '-../
,.---- xres
2
2
Y "1
Yl
11 "
(a)
(b)
Yo
r--...
Yl
II
~.~
><
rL_
f
Y4tol
Y2
I II
Y3
I III
11.5
8>0
f15
1.5
xo
xO
0---
(c)
IXI
2X2
3X3
(d)
Figure 6.5: Circuit implementation (a) Quaternary 2 to 1 multiplexer; (b) Quaternary reshaper;
(c) Threshold function generator; (d) Quaternary 4 to 1 multiplexer

The performance is measured when output is loaded with C
101
O.lpicofarad. The circuit
with heavy capacitive load is to estimate its driving capabilities. The clock-to-Q time is measured
from the 50% point of the clock logic-level transition of the associated output transition. The
clock-to-Q time for the best-case output transition from logical-One to logical-Zero is 0.15ns
while the time for the worst-cased from logical-'IWo to logical-One is 0.59ns. The rising time
and falling time are measured from the 10% point to 90% point of the transition level, which
borrows from the conventional definition in the binary logic measurement. It can be seen that
the falling time is much longer than the rising time for this design. Table 6.2 shows the average
value.
Table 6.2: Performance measurement of the MVFF (ns)
Transition of logic levels Clock-to-Q time Rising time Falling time
0.14
0++1
0.25
0.34
0.42
0.05
0.14
1++2
0.29
0.06
0.06
2++3
0.24
0++2
0.28
0.17
0.27
0.08
0.25
1++3
0.08
0.11
0.33
0++3
5 .au
-.----------.,--.
--r----r - - - - - - - - - ...,.--....,
CP
au
----------1
'- - - - - - - - - -
5 .au -.----------- ,
----------------T"",- - - - - ,
I
I
I
I
o
au
S.1U
1 - - - - - - - - r - - - - - - - -I
1- r - - - - - - - - - - - -
-I
r--------------------~-----------------I
-e.1u
I-
es
..
zzzL
1------
zens
'
------r --
-iens
Time
Figure 6.6: PSpice simulation of the proposed flip-flop
_____ 1
Gens
102
A comparison for the storage capacity between the MVCP controlled quaternary D flip-flop
in Fig. 6.3 and the two-digit shift register in Fig.6.2 (b) can be made. They use the same number
of latches. However, the former can output a four digit quaternary signal in one clock cycle while
the latter can only output a two digit quaternary signal in two clock cycles. The reason is that
the proposed flip-flop can receive input signal when the quaternary CP takes any of its four
values while the latter can only do so when CP = 3.
Average power dissipation is measured with 5V and 50MHz clock for the proposed circuit and
evaluated for a time period of 200ns, which is 1.49mW. For the fair comparison of the proposed
circuit and that in Fig. 6.2, the same PSpice simulator and 0.5um technology are used to simulate
the circuit in Fig. 6.2. The power dissipation is 2.56mW with 5V and 200MHz clock. The reason
of using 200MHz clock is that BCP based flip-flop samples the input date only when CP
and the same data throughput as that of the proposed circuit is required. The power reduction
is 41.3% with the area trade of 29.8%.
Table 6.3: Area and Power dissipation for the proposed circuit and the circuit in Fig. 6.2(b)
No of Transistors Power dissipation (mW)
2.56
Circuit in Fig. 6.2(b)
80
1.49
114
Proposed circuit
-29.8
41.8
Improve (%L
- -
6.5
Summary
In this chapter, the present BCP controlled multiple valued master slave flip-flops in series are
analyzed first. Then MVCP controlled multiple valued flip-flop in parallel is proposed. Based on
transmission switch theory, the CMOS circuit of the proposed flip-flop is implemented. PSpice
simulation shows that it has the correct operation. Although only the quaternary D flip-flop is
designed, the result can be extended to the design of ternary and higher radix flip-flops. This
design changes the traditional design method of multiple valued flip-flops. It uses each level of
a quaternary clock signal to receive an input signal, which makes the logic structure of multiple
valued flip-flops more flexible. The proposed flip-flop has the following characteristics:
1. Employing multiple valued clock signal ensures the consistency of processing signals for
103
multiple valued flip-flops, which eliminates the requirement to provide the same circuit with both
multiple valued signals and binary signals.
2. An n-valued CP controlled n-valued flip-flop can shift out an n-digit n-value signal during
one upward clock cycle while a BCP controlled multiple valued flip-flop can only shift out onedigit n-valued signal during one binary clock cycle. The proposed circuit is characterized by
reduced power dissipation.
3. Because there are possibly multiple connections between each latch and the output multiplexer, various logic functions can be obtained by changing the connections, which adds flexibility
to the design of the flip-flops.
Chapter 7
Low power XOR gate decomposition

7.1
Introduction
With the rapid increase in the density and the size of chips and systems, power dissipation
is becoming critical concern in VLSI design. Since CMOS technology is predominant in the
realization of today's ICs and CMOS devices are intrinsically low-power consuming, CMOS
has become the reference technology. Power dissipation in digital CMOS circuits is dominated
by dynamic dissipation, which results mainly from the charging and discharging of the node
capacitances [89]. It can be modeled as in equation 7.1
Power
1 2
'
" CiSAi
= 'iVddfdk
L.J
(7.1)
where
Vdd
is the supply voltage and
fdk
is the clock frequency which are determined by the
technology, Ci is the physical capacitance at the output of the node (i is the node number) and
SA i (referred to as the switching activity) is the expected number of output transitions per clock
cycle[131]. The summation is taken over all nodes of the logic circuit. As is known, Ci and SAi
can be optimized during the design process.
Logic functions can be expressed canonically based on AND JOR (NAND & NOR) operators.
Extensive research has been done on developing low-power techniques in AND JOR or NAND
104
CHAPTER 7. LOW POWER XOR GATE DECOMPOSITION
105
& NOR based circuits [91]. However, in certain applications, XOR realizations have attractive
advantages over the conventional AND JOR NAND&NOR logic especially for functions which do
not produce optimization solutions based on these operators. The XOR circuit is easier to test
and may require fewer gates and interconnections [3]. Unfortunately, the techniques for synthesis
and optimization of logic based on Reed-Muller are much more difficult than those based on the
standard Boolean expressions.
With recent development and increasing use of various field programmable gate array (FPGA)
devices[21] where the XOR gate is already manufactured as a basic component, synthesis and
optimization of XOR gate based circuits are receiving much more attention than before. More
recently, there has been some success in achieving area reduction by employing optimization
techniques specifically targeted towards initial AND jXOR representations in the well known
Fixed Polarity Reed Muller ( FPRM) form [114][107]. The low power optimization of XOR gate
based circuits is still in its infancy.
In technology mapping, the subject net list is usually first decomposed into a netlist composed
of only inverters and two-input XOR gates. The decomposition can have a significant impact on
the power dissipation of the final implementation [74][131]. Looking at the existing work done
by Narayanan and Liu [74] and Zhou and Wong [131]' the former is for static logic and the latter
is for dynamic logic. However, as we will show later, the algorithms in [74][131] are not optimal.
In this chapter, we will deal with the low power XOR gate decomposition problem.
The rest of this chapter is organized as follows. In section 7.2, some definitions and terminology are introduced. In Section 7.3, some previous work is reviewed. Section 7.4 presents a
new algorithm for low power XOR gate decomposition while Section 7.5 shows the experimental
results. Finally, a brief summary is given in Section 7.6.
7.2
Definitions and terminology
To implement a complex function involving many inputs, a form of decomposition into smaller
subfunctions is required such that the subfunctions fit into the primitive elements to be used in
the implementation.
106
Definition 7.1. Low power XOR gate decomposition: a multi input XOR gate is decomposed
into a tree of two-input XOR gates, which is of optimal power dissipation.
A tree is defined by the following definition.
Definition 7.2. Let T = (V,E) represent a decomposition tree, and Pr(v), for any v E V,
denotes the output signal probability of node v. Each node has two children. The primary
inputs are called leaves of the tree and the primary output is called the root of the tree.
Each primary input signal into a tree is treated as a random variable and its probability is
defined as follows.
Definition 7.3. Signal probability for signal x is defined as the probability of x being 1.
To estimate the power dissipation of a tree, according to equation 7.1, Ci and SAi need to
be known. Each node in the tree corresponds to a two-input XOR gate. Suppose Ci is constant
for each two-input XOR gate. Then, the only parameter we need to compute is SA i .
Under the assumptions of zero delay model with temporal and spatial independence [74][131],
given input signal probabilities and a decomposition tree, the probabilities of internal signals can
be computed according to the following definition of XOR operator.
Definition 7.4. XOR operator: f(Xl, xo) =
Xl
El7 Xo = XlXO
+ XlXO
if
Xl
and Xo are two input
variables. Hence, the probability of f(Xl, xo) is as shown in equation 7.2 [74][131]:
PrJ
= Pr(Pr,xl ,Pr,xQ) = Pr,xl + Pr,xQ
- 2Pr,X1Pr,XQ
(7.2)
The calculation of the corresponding switching activity depends on whether the circuit is
implemented static logic (which does not take into account the timing behavior and is strictly a
function of the topology and the signal statistics) or in dynamic logic (which takes into account
the timing behavior of the circuit).
For static logic, the transition probability assuming independent inputs is the probability
that the output will be in the zero state multiplied by the probability that the output will be in
107
the one state. For example, if the signal probability being zero is Po and Pi represents the signal
probability being one, the transition probability from zero to one is prO -+ 1]
= POPi = po(1- po).
In general, for a digital signal, pro -+ 1] is equal to p[1 -+ 0]. Therefore, the switching activity
for gate i can be expressed as 7.3 [131]
SAi = Pr,i[O -+ 1]
+ Pr,i[1 -+ 0] =
(7.3)
2Pr,i(1 - Pr,i)
The main difference between static logic and dynamic logic is that dynamic CMOS works
using a precharging circuit. The output of a circuit can be precharged high or low then the
resulting value is evaluated in each clock cycle. Therefore, the output transition probability does
not depend on the sate of the inputs but rather on just the signal probabilities[24]. There are
two implementation models, which are precharged to 0 or 1 and evaluated to the result value.
Take the former for example, that is, a gate output is first precharged to 0 at the leading edge
of the clock signal and then evaluated to the result value at the falling edge. Therefore, if the
output probability is Pr,i, the switching activity for gate i will be 2Pr,i where one Pr,i comes
from the precharging and one Pr,i comes from the evaluation.
Then, for dynamic logic, the switching activity for gate i is given by equation 7.4 [89]
SA i
= 2Pr,i
(7.4)
Fig. 7.1 shows the switching activity as a function of signal probability both in dynamic logic
and static logic. It can be seen that if only considered in term of switching activity, static logic
is always better than dynamic logic. If the two input signal probabilities of an XOR gate are
Pr,xQ
= 0 and Pr,xl = 1 respectively, then the output signal probability of XOR gate is
=1
according to equation 7.2. One transition, evaluated to 1 at the falling edge of the clock, is
followed by another transition, precharged to 0 at the leading edge of the clock. Hence, in this
particular case, the switching activity of an XOR gate is equal to that of a clock signal, which is 2.
For static logic, the maximum switching activity is 0.5 when signal probability is 0.5. However,
108
in terms of power dissipation, the same conclusion can not be deduced from switching activity
since dynamic logic usually has smaller load capacitance.
2
>-.
.t:!
.~
1j
ro
,rY
1.5
c:tynamic
.., .... ....
0)
.... ....
..,. ,.
,. .,.
static
:. 0.5
..kl
0.5
Signal probability
Figure 7.1: Switching activity versus signal probability

Given an n-input XOR gate with primary input signals I
{Xn-l, Xn-2,"
,xo} and cor-
responding signal probabilities {Pr,x n- II Pr,x n-2' ... ,Pr,xQ}' based on the above discussion, the
procedure for the power estimation of the XOR gate decomposition is as follows.
1. Given the primary input signal probabilities, construct a tree of two input XOR gates.
2. Compute the output probability for each gate.
3. Compute the node switching activity using equations 7.4 or 7.3.
4. Compute the total switching activity using equation 7.1.
Then, the problem for low power XOR gate decomposition can be described as follows.
Low power XOR gate decomposition: Given an n-input XOR gate with input signals
1=
{Xn-l, X n -2,'"
,xo} and corresponding signal probabilities {Pr,Xn-llPr,Xn-2"" ,Pr,xQ}' two
input XOR gates are used to construct a tree T
(V,E) with Pr,Xn-llPr,x n-2'''' ,Pr,xQ as
its leaves such that equations 7.5 and 7.6 are minimized for dynamic logic and static logic
respectively.
For dynamic logic,
SA
2::
109
2Pr,i
(7.5)
2Pr,i (1 - Pr,d
(7.6)
iEV-J
For static logic,
SA =
2::
iEV-J
Where SA is the total switching activity and Pr,i is the signal probability of an internal node.
7.3
Previous work
In [131], the authors analyzed some optimal properties and proposed an algorithm to solve the
problem of lower power XOR gate decomposition in dynamic logic.
Algorithm 7.1. Zhou - Wong's algorithm
First, sort the input probabilities which are greater than 0.5.
Then, combine them pairwisely from the largest to the smallest.
Finally, iteratively combine the two signals with the smallest probabilities until there is one
signal.
Based on the analysis of XOR operation and assuming that the signals and their complements
are available, Naraynana and Liu [74] presented an algorithm as follows.
Algorithm 7.2. Naraynana - Liu's algorithm
======================================================================================================
Let k be the number of primary inputs to the XOR tree with probabilities greater than 0.5.
If k is even, replace all k inputs by their respective complements.
If k is odd, replace any k - 1 inputs by their respective complements.
110
Then simply use the Huffman algorithms to build the optimal tree.
Take a simple example to show how their algorithms work in Fig. 7.2. The problem has
three inputs and corresponding signal probabilities are 0.1, 0.45 and 0.9 respectively. Both
algorithms share the same decomposition in Fig. 7.2(b). The total switching activities are 1.984
from equation 7.5. However, if the assumption that both the signals and their complements are
available is preserved, the above decomposition is not optimal. Suppose signal
X2
and
Xl
are
replaced by their complements and it can be proved that the functionality is still preserved. Then,
there is another decomposition as shown in Fig. 7.2(c) with a total switching activity of 1.424,
which is smaller. It is realistic to assume that both signals and their complements are available
like FPGA mapping. Hence, in the chapter, all discussions are based on this assumption.
x 2 0.9
xl 0.45
Xo 0.1
Xo 0.1 ~.46
Xo 0.1~.18
xl 0.45
x2 0.9
x2 0.1
xl 0.5
0.532
(b)
(a)
0.532
(c)
Figure 7.2: (a) Multi input XOR gate; (b) Published decomposition in [131] and [74]; (c) A new
decomposition
Actually, both the Naraynana-Liu's algorithm and the new decomposition in Fig. 7.2(c)
use the polarity transformation. However, any n-variable Boolean function can be expressed
canonically by the Fixed Polarity Reed-Muller (FPRM) form as follows:
2n-l
!(Xn-l, Xn-2, ...
,xo) = E!1
bi'rri
(7.7)
i=O
The
1r-
xn -ti:n-2' .. xo, the subscript i can be written as a

io), bi E {O, 1}, x is a literal which can be X or X,
terms can be represented as
binary n-tuple i
= (in-lin-2'"
1ri
111
1, ij
Xj = {
and 0
=0
(7.8)
Xj, ij = 1
n - 1.
An FPRM form of a Boolean function can be identified by a polarity p, P = (Pn-IPn-2'" po).

Pj E {O,l}. Pj
= 0 if Xj is true otherwise Pj =
1. For FPRM forms, every variable can only be
either true or complemented but not both. Therefore, there are 2n polarities for an n-variable
function.
Definition 7.5. If P = (Pn-IPn-2'" po) and q = (Qn-1Qn-2'" qo) are the two polarities of an
n variable function under FPRM forms. If the number of binary bits that are 1s is even, it is
called even polarity, otherwise it is called odd polarity. If all variables appeared in equation 7.7
are in true forms, the polarity of expansion is known as positive polarity or polarity zero.
Obviously, an n variable XOR function is a special case of an n variable FPRM form where
the number of 1f-terms equals to the number of variables and 1fi
Xi
and its polarity is zero.
For example, an n-variable XOR function under polarity zero can be expressed as
!(Xn-l, Xn-2,'"
7.4
,xo)
Xo EI1 Xl EI1 X2'" EI1 Xn-l
(7.9)
Proposed low power XOR gate decomposition
Dynamic logic is used to illustrate the technique. For dynamic logic, according to equation 7.5,
minimizing the total switching activity can be achieved by minimizing the total internal signal
probabilities.
In terms of polarities, the algorithm proposed in [131] is optimal only under polarity zero.
Although the authors in [74] tried to use polarity transform to optimize power dissipation, they
only tried one of 2n polarities, because, for given input signal probabilities the selected polarity is
fixed according to their algorithm. For example, for a three input problem shown in Fig. 7.2, the
112
selected polarity is zero. For a five input problem whose input signal probabilities are 0.1, 0.2,
0.55, 0.7 and 0.9, the selected polarity using the algorithm in [74] is 3 or 5. As we know, different
polarities have different function forms, which result in different power dissipation as shown in
Fig. 7.2. Hence, if 2n polarities are searched, the best polarity for low power decomposition can
be found. However, for reasons that will become apparent in Section 7.4.1, the search space can
be reduced to 2n -
7.4.1
1 -
1.
Theoretical results
For an n variable function, there are 2n polarities and correspondingly there are 2n FPRM forms.
FPRM forms can be transformed from one to the other with polarity transformations [114]. From
definition 7.5 and starting from positive polarity, all polarity transformations can be categorized
into two classes: Odd Polarity Transformation (OPT) and Even Polarity Transformation (EPT).
Lemma 7.1. Suppose there is an XOR function consisting of n inputs based on equation 7.9. If
an OPT is selected, one more

1r -
1r -
term,l, is added while if an EPT is selected, the number of
terms remains the same.

This is trivial because
x$1 = x and x$ x = O.
Take the three inputs in Fig. 7.2 for example.
If transformations of polarities 2 (010), odd polarity, and 6 (110), even polarity, are derived from
the positive polarity form, the function forms are shown as follows:
f(x2,xI,xo) = 1$XO$XI$X2
(7.10)
f(X2' Xl, xo) = Xo $ Xl $ X2
(7.11)
Form Lemma 7.1, there is a following corollary.

Corollary 1. For any XOR tree T consisting of n inputs, the OPT requires one additional input
node compared to the EPT.
113
From Lemma 7.1, Corollary 1 is obvious because for a tree of two input XOR gates, the
additional input needs one more two-input XOR gate. Hence, EPT can preserve the size of an
XOR tree between polarity transformations.
Theorem 7.1. Given an n variable junction, there are one zero polarity, 2n-l_1 even polarities
and 2n -
odd polarities [126}.
This can be proved by induction on n as below.

(1) When n is 1, there are two polarities 0 and 1. Obviously, it satisfies the Theorem 7.1.
(2) Suppose when n is k, k > 1, polarity p' = (Pk-1Pk-2'" po) and there are one polarity
zero, 2k- 1 - 1 even polarities and 2k- 1 odd polarities and Theorem 7.1 is true. When n is k + 1,
polarity p
(PkPk-l' .. po).
(a) When Pk is 0, the distribution of polarities is the same as the
case n = k. Hence, there is one polarity zero, 2k (b) When Pk is 1, then P = 2k
+ p'
where
"+"
1-
1 even polarities and 2k -
odd polarities.
is the arithmetic addition. Compared to the p',
the polarity distribution changes as follows: Polarity zero::::}odd polarity; 2k-l -1 even polarities
::::}2k-l_1 odd polarities; 2 k- 1 odd polarities::::} 2k- 1 even polarities. Hence, there are 2k- 1 even
polarities and 2k -
odd polarities. From both (a) and (b), Theorem 7.1 is true when n is k
+ 1.
Corollary 1 and Theorem 7.1 tell us that considering area constraints, EPT is preferred and
the best polarity could be found by searching 2n -
1 -
1 even polarities.
For any specified polarity, all input signal probabilities can be deduced from the input signal
probabilities under polarity zero. However, given a set of input signal probabilities, it must be
one of following three cases.
Case 1: At most one input signal probability is greater than 0.5.
Case 2: All input signal probabilities are greater than 0.5.
Case 3: Some are greater than 0.5 and some smaller than 0.5.
For each case, the optimal solution for low power XOR gate decomposition is to find out
an optimal combination of input signals with the merging function in equation 7.2 such that
the total internal probability is minimized. For Cases 2 and 3, the theorem in [131] is used to
combine input signal probabilities greater than 0.5 pairwisely.
114
Theorem 7.2. (Zhou and Wong) In an optimal XOR gate decomposition, all inputs of probabilities greater than 0.5 are combined pairwisely from largest to smallest.
Then, using the above theorem, Cases 2 and 3 can be transformed into Case 1. For Case 1,
it is shown that under some conditions a variation of Huffman's algorithm, which is a little bit
different from the original one, can give an optimal solution[74][131]. It is defined as follows.
Definition 7.6. Mini-Huffman algorithm: [107]Start from all the input signals; combine the
two signals of minimum probabilities and substitute them with a new signal; continue the process
until there remains only one signal.
7.4.2
Proposed algorithm
Based on the above properties, a new algorithm to solve the low power XOR gate decomposition
problem is proposed as follows. First, according to input signal probabilities given with a specified
polarity, an optimal tree for low power XOR gate is derived based on Theorem 7.2 and MiniHuffman algorithm, as described in Procedure 7.1 and corresponding pseudo-code in Algorithm
7.3.
Procedure 7.1. Given the input signal probabilities set for an n variable XOR function with
polarity p, carry out Steps 1 to 3:
Step 1. If the problem belongs to Case 1: At most one input signal probability is greater than
0.5, use mini-Huffman algorithm to combine the signal probabilities and construct a tree and
compute the total switching activity;
Step 2. If the problem belongs to Case 2 or Case 3, according to Theorem 7.2, combine the
signal probabilities using equation 7.2 until Case 2 or Case 3 is changed into Case 1. Then, go
to Step 1;
Step 3. Output the tree and its total switching activity.
Then, other polarities are calculated and their corresponding switching activities are compared to find the best polarity which has the minimum switching activity. However, based on
Theorem 7.1, only 2n -
1 -
1 even polarities need to be computed. This is shown in Procedure
7.2 and corresponding pseudo-code is given in Algorithm 7.4.
OHAPTER 7. LOW POWER XOR GATE DEOOMPOSITION
115
Procedure 7.2. Given the input signal probabilities set for an n variable XOR function with
polarity zero, set bestPolarity, SAp and OptSA to be 0, respectively. For any p, O<p$. 2n - 1,
carry out Steps 1 to 5.
Step 1. Use Procedure 7.1 to compute switching activity (SAo) with polarity zero. Let OptSA
SA o ,Step 2. Let p
= p+1,-
Step 3. If p is OPT, go to Step 2. If p is EPT, use Procedure 1.1 to compute SAp with
polarity p.
4.
Step
If SAp is less than OptSA, let OptSA=SAp and bestpolarity=p. If p$. 2n - 1,go to
Step 2. Otherwise, go to Step
4.
Step 5. Output OptSA and bestPolarity.
Algorithm 7.3. Pseudo-code of Procedure 1.1

~====================================================================================================
Input: a set of signals X

(
"
= {Xn-l, Xn-2,'"
Pr,xn-l' Pr,x n-2' ... ,Pr,xoJ under polarity p
Output: a decomposition tree T

Decomp(X)
{
if /X/ ==1
I
return Pr,xn-l'
else
(
if(Case 1)
(
T
= mini-Huf fman(X),-
return T,-
J
if(Case 2 or Case 3)
,xo} and corresponding probabilities
116
S = {Si = combine{xn-i,Xn-i-t}},{XI E XI:S; X); //XI is a set of primary

input signal probabilities greater than 0.5
T
= minLHuf fman{X + S
- Xd;
return (T + S);
}
}
Algorithm 7.4. Pseudo-code of Procedure 7.2
Input: a set of signals X
{Xn-l, Xn-2,'"
xo} and { Pr,x n -llPr,x n - 2 " " ,Pr,xo} under
polarity zero
Output: OptSA and bestPolarity / /OptSA-the smallest switching activity for a tree; bestPolaritythe polarity corresponding
/* Decompose
under zero polarity*/
bestPolSearch(X)
{
if /X/==l
OptSA
= Pr,xn-l;
else
{
OptSA=Decomp(XY.SA; / /SA - switching activity
}
bestPolarity=O; / /bestPolarity - the polarity corresponding OptSA is named bestPolarity
polarityN0 = 1 < < n; / /The number of polarities for an n - input variable function is named
polarityNo
for (p = 1; p
< polarityN0; p + +)
(
if(p%2 == 1);
117
else
(
/ /decomp
SAp
/X/ under polarity p
= Decomp(X) . SA; / /SA p- switching activity under polarity p
}
if( SAp
< OptSA)
{
OptSA = SAp;
bestPolarity
= p;
}
}
Output OptSA and bestPolarity;
7.5
The proposed algorithm is implemented in C and compiled by the GNU C compiler. The time
complexity for the mini-Huffman procedure is O(nlogn)[131]. The proposed algorithm has a time
complexity of O(2nnlogn) where n is the number of variables. In order to compare our results
with those from other algorithms, Zhou-Wong's algorithm and Narayanan-Liu's algorithm are
also implemented. All the three algorithms are tested on a personal computer with PIlI 550
CPU and 64M RAM under Linux operating system. Two sets of experiments have been carried
out.
First, given a set of input signal probabilities, the best polarity (the lowest switching activity)
is found and its corresponding switching activity is computed and compared to that under
polarity zero to see the efficiency. Table 7.1 shows the experimental results for input size from 3
to 13 variables. Thirteen input signal probabilities are generated under randomly selected seed
number 500, which are 0.98, 0.57, 0.17, 0.70, 0.71, 0.18, 0.03, 0.14, 0.69, 0.70, 0.81, 0.32, 0.63,
118
respectively. For different input sizes, the input signal probabilities are chosen from left to right
in the above list. For example, for three inputs, the signal probabilities are 0.98, 0.57 and 0.17
while for four inputs they are 0.98, 0.57, 0.17, 0.70, and so on. In Table 7.1, Column 2 from
left shows the best polarity and its corresponding switching activity SAp, Column 3 shows the
switching activity SAo decomposed under polarity zero and last column shows the improvement
rate which is defined as follows:
Improv
SAo - SAp
eJ!
-p
(7.12)
Table 7.1: Best polarity and switching activity for low power decomposition
INs
Best Decomp
Polarity Zero Improv
Best Polarity
SAp
SAo
(%)
0.6388
0.8884
39.07
3
3
4
1.0742
11.62
1.1990
9
1.4895
1.6299
9.43
5
27
1.9693
6.10
6
27
1.8560
2.1771
12.23
1.9399
7
3
2.2620
2.5049
10.74
8
3
2.7286
2.9525
8.21
9
257
3.1781
3.4198
7.61
3
10
3.5494
11
3.6889
3.93
305
12
4.0205
4.1610
3.49
305
4.4925
4.6424
3.34
13
2355
Decomposed with the best polarity, it can be seen that the improvement could be as high as
39% compared to the result with polarity zero for this randomly selected instance.
Second, for each input size from 3 to 13 variables, 100 sets of input probabilities are randomly
generated and then run on each of the above algorithms. For each instance X, let Opt(X)
represents the optimal solution, which has the lowest switching activity from the three algorithms.
The following parameter defined in equation 7.13 is used to measure the performance of solution
S(X).
= S(X) - Opt (X) o/c

Opt(X)
119
(7.13)
From the definition in 7.13, obviously, the higher the rate R is, the worse the performance
of the algorithm will be. For each algorithm, among all 100 problems of the same input size,
the number of non-optimal solutions, the maximum ratio and the average ratio are computed.
They are represented by N on Opt , M R and AR, respectively. M R is to measure the worst
performance among all 100 problems of the same input size while AR is to measure the average
poor performance. The average CPU time is also reported in Table 7.2.
Table 7.2: Experimental results for the proposed algorithm and published algorithms
INs Narayanan-Liu [74]
Proposed
Zhou-Wong [131]
Non MR
AR Non MR
AR Non time
Opt
Opt
(%) (%)
(%) (%) Opt 10- 2 8
41
46.04 7.79
3
50
69.77 10.87
0
0
4
40
44.41 8.11
40.58 8.88
0
0.1
72
44
5
35.75 8.57
89
28.39 8.28
0
0.3
6
53
34.43 9.08
92
28.18 7.42
0
0.8
51
24.10 5.59
25.04 6.76
92
0
1.0
7
45
20.01 4.91
8
27.43 6.31
0
1.9
95
48
22.50 5.90
14.47 4.38
4.2
98
0
9
10
18.31 6.80
7.25
4.00
58
98
9.7
0
11
5.28
3.63
21.8
56
15.17 6.09
99
0
12
54
48.6
13.37 5.88
4.79
3.56
0
99
13
47
4.06
3.27
0
108.0
11.77 5.08 100
From the results reported in Table 7.2, the number of non-optimal solutions for the proposed
algorithm is always zero in comparison with Narayanan-Liu's and Zhou-Wong's algorithms. In
other words, the switching activity from the proposed algorithm is always smaller than that from
Narayanan-Liu's or Zhou-Wong's algorithm. It can be seen that the average poor performance
is decreased with the increase of input size for Zhou-Wong's algorithm but this does not happen
for Narayanan-Liu's algorithm. For smaller input size like 3 and 4 variables, Narayanan-Liu's
algorithm is better than Zhou-Wong's algorithm. The reason is that in this case the number
of polarities is smaller and there is a higher chance for Narayanan-Liu's algorithm to find a
120
good polarity than Zhou-Wong's algorithm. For example, for three inputs: 0.10,0.80,0.90, using
Narayanan-Liu's algorithm the polarity is 3 and the switching activity is 0.488 while using ZhouWong's algorithm it is 0.568 under polarity zero.
7.6
Summary
Using the polarity transform under the assumption that both signals and their complements are
available, a new solution to decompose low power XOR circuits is proposed. The algorithm is
implemented in C. The time complexity of the algorithm is O(2nnlogn). The improvement of
switching activity is significant compared to published results. It should be pointed out that the
algorithm can be applied to static logic as long as the calculation formula of switching activity
in equation 7.5 is substituted by the static logic formula in equation 7.6.
Chapter 8
Power and area optimization of FPRM

functions
In the past, the task of VLSI designers has been to explore the area-time implementation space.
In recent years, power is being given comparable weight to area and time [28].
Up to now, most researchers have focused on developing optimization techniques for area
and power dissipation in AND/OR or NAND & NOR based circuits [91]. However, in certain
applications, XOR realizations have advantages over the conventional AND/OR or NAND &
NOR logic. XOR circuits are easier to test and may require fewer gates and interconnections
[3][87]. Applications of Reed-Muller circuits have so far not become popular due to the following
two obstacles.
1. XOR gates used to have slow speed and require large silicon area to realize in comparison
with OR gates.
2. The problem of optimization of Reed-Muller functions is difficult although there has been
a great deal of research in recent years [64][114].
With the development of new technologies and the advent of various field programmable
gate array (FPGA) devices [21], the first obstacle has been solved. In programmable devices, the
XOR gate is either easily configured in "universal modules" or directly available. For example,
logic blocks can be configured as various two-input gates such as XORs, ANDs and NANDs in
the AT6000 FPGA series from ATMEL Corporation while in other FPGAs, both AND and XOR
121
CHAPTER 8. POWER AND AREA OPTIMIZATION OF FPRM FUNCTIONS
122
gates are available in the macrocells or logic array blocks [76]. For the second obstacle, more
recently, there has been some success in achieving area reduction by employing optimization
techniques specifically targeted towards initial AND /XOR representations in the well known
Fixed Polarity Reed Muller expansion [64][114][107].
Usually, area synthesis of circuits promises more global power savings. However, power dissipation for a circuit depends on input pattern, which requires more specific low power techniques.
For low power optimization of Fixed Polarity Reed-Muller (FPRM) functions, to the best of our
knowledge, the only results in the literature are algorithms proposed by Zhou and Wong [131]
and Narayanan and Liu [74]. The former is for low power XOR gate decomposition, which is the
special case of FPRM functions and hence can not solve low power problem for general FPRM
functions. And the latter is for the FPRM functions with the implementation of static logic.
However, as will be shown later, the algorithm in [74] is not optimal and can only be applied to
some special FPRM functions.
The objective of area and power optimization is twofold. This chapter proposes a frame to
estimate power dissipation based on FPRM functions, discusses the effect of polarity on area
and power dissipation and introduces an optimal algorithm for area and power minimization of
single output FPRM functions. The rest of the paper is organized as follows. In Section 8.1,
some definitions and terminology are introduced. Some previous work is reviewed in Section 8.2.
Section 8.3 presents a frame of power estimation and a cost function to estimate area and power
while an optimisation algorithm is proposed in Section 8.4. Finally, some experimental results
are shown in Section 8.5 and summary in Section 8.6.
8.1
Definition and terminology
For convenience, some definitions are restated here.

Any n-variable Boolean function can be expressed canonically by the exclusive SOP form as
follows.
123
2n-1
f(xn-l, X n-2,'" ,xo)
= EB
l: b(lri
(8.1)
i=O
where the subscript i can be expressed into a binary form as i

the 1r-terms can be represented as 1ri
= Xn -IX n -2 . "
X.
J
={
1,
.
Xj,
(in-lin-2'" io), bi E {O, 1},
xo, x is a literal which can be x or X,
ij = 0
ij
(8.2)
=1
In equation 8.2, j is from 0 to n - 1. The form in equation 8.1 is also known as the positive
polarity Reed-Muller (PPRM) form if all the variables are in true forms. PPRM forms can be
extended to FPRM forms with any fixed polarity p, p
(Pn-IPn-2'" po), where variables can
only be either true or complemented, but not both. If a binary bit of p,
Pj ,
is 0 (or 1) then the
corresponding variable is in the true (or complemented) form. Hence, there are 2n polarities for
an n-variable function. Each FPRM form can be identified by its polarity.
Example 8.1. A four variable function under polarity 0 is given:
X2XI
EB X2XIXO EB
X3
EB
X3 X O EB X3X2
EB
X3X2XO
f(X3, X2,
Xl, Xo) = XOEBXIXOEB
EB X3X2XI
Then, under polarity 5, it has
f(X3, X2,
Xl, Xo)
= 1 EB Xo EB Xl EB EBX2XIXO EB X3XI EB X3X2XO EB X3X2XI
(8.3)
while under polarity 14, it will be
f(X3, X2,
Xl, Xo)
Xo
EB
X2
EB X2XIXO EB
X3
EB X3XI EB X3X2XO EB
X3X2XI
(8.4)
The above functions can be implemented using AND gates and XOR gates with suitable
124
technology. XOR gate has defined in Definition 7.4 while two input AND gate is based on the
following definition of AND operator.
Definition 8.1. For two-input AND operator: !(Xl,XO)
= Xl' Xo , the probability of !(Xl,XO)
is [89]:
PrJ
= Pr{Pr,xllPr,xo) = Pr,xl
. Pr,xo
(8.5)
Here, Pr,xi is the input signal probability, which was defined in Definition 7.3 while PrJ
represents the output signal probability of the corresponding gate.
8.2
Related work
The total SA of a specific function form depends on how to combine the input signal probabilities,
which refers to low power gate decomposition. The problem for gate decomposition has been
extensively studied [107][74][131].
In [107], AND gate decomposition was discussed. For dynamic logic, it is found that a modified Huffman's algorithm gives optimal solutions. The modified version of Huffman's algorithm
is called Mini-Huffman algorithm which was described in Algorithm 7.6.
Take a simple example to explain the algorithm. For three inputs (X2, Xl, XO) and corresponding signal probabilities are 0.9, 0.45 and 0.1 respectively, Fig. 8.1{a) shows the above algorithm
applied to decompose a three input AND gate into a tree of two-input gates in dynamic logic.
The switching activity is 0.171.
In [131], the authors analyzed some optimal properties and proposed an algorithm to solve
the problem of lower power XOR gate decomposition in dynamic logic in Algorithm 7.1.
Fig. 8.1{b) shows the decomposition based on algorithm 7.1. The switching activity is 1.984.
It can be seen that this algorithm only works for XOR gate decomposition.
Based on the analysis of XOR operation and assuming that both the signals and their com-
125
plements are available, Naraynana and Liu claimed that they use the selection of polarity vectors
to solve low power logic synthesis for XOR based circuits in Algorithm 7.2.
However, this algorithm can not handle the reduction of tree size and only works for some
special cases of FPRM functions. For example, consider two five variable problems, which share
the same input signal probabilities: {Pr,x4,Pr,x3,Pr,x2,Pr,Xt,Pr,xo} = {0.1, 0.3, 0.7, 0.8, 0.9}. Tho
function forms are shown in equations 8.6 and 8.7:
f(X4, X3, X2, Xl, Xo)
f(X4, X3, X2, Xl, XO)
= XIXO E9 X2XO E9 X4 E9 X4X3XO
XIXO
(8.6)
(8.7)
E9 X2XO E9 X4Xl E9 X4X3XO
Obviously, algorithm 7.2 works for the function in equation 8.6. If Xl and
X2
in equation 8.6
are replaced by their complements, then an alternative representation of the function is:
f(X4, X3, X2,
Xl, XO)
= XIXO E9 X2 X O E9 X4 E9 X4X3XO
(8.8)
which has the same size and the same logic functionality as in equation 8.6. However, if the
same algorithm is applied to the function in equation 8.7, the function will be changed into:
Xo
XI
0.1
x2
0.9
0.45~
J1
Xo
0.1
x I 0.45
LJ
(a)
X2 0.9
(b)
Figure 8.1: (a) Multi input AND gate decomposition in [107]; (b) Multi input XOR gate decomposition in [131]
f(X4, X3, X2, Xl, XO)
= X1XO EEl X2XO EEl X4Xl EEl X4X3XO
126
(8.9)
which offers different functionality from the function in equation 8.7. This is easily tested by
applying the formula X = 1 EEl x to
X2
and
Xl
X1XO
f(X4, X3, X2, Xl, XO)
in equation 8.7, which will result in
EEl
X2XO
EEl X4 EEl
X4 X l
EEl
X4X3XO
(8.10)
Obviously, equation 8.10 is different from equation 8.9. Hence, the algorithm does not work
well with the function in equation 8.7.
Based on Algorithm 7.2, the algorithm only tried one of 2n polarities because given input
signal probabilities the selected polarity is fixed. For example, for a three input problem shown
in Fig. 8.1, the selected polarity is zero while for a five input problem as the above, the selected
polarity is 7. Therefore, the selected polarity may not be optimal.
Hence, an FPRM function with multi variables can be implemented using trees of two-input
AND gates and a tree of two-input XOR gates. Given the input signal probability set, the
switching activity can be computed using the above algorithms.
8.3
Polarity based area and power minimization of FPRM functions
8.3.1
Power estimation
As stated in Section 7.1, all discussions here refer to CMOS technology. Consider dynamic power
in CMOS circuit. The power can be modeled as in equation 7.l.
Power dissipation for a circuit is an input pattern involved problem. Any FPRM form can
be implemented by AND gates and XOR gates. Before technology mapping, conventionally, all
multi input AND gates and XOR gates need to be decomposed into a tree of two input gates,
127
which is called AND & XOR tree. A tree is defined in Definition 7.2.
Furthermore, each primary input signal into a tree is treated as a random variable and its
probability is defined as in Definition 7.3.
Power estimation of an AND & XOR tree is similar to that of an XOR tree in Section 7.2.
The difference is that there are two kinds of nodes in the AND & XOR tree, namely XOR gate
output and AND gate output.
Given an n-input FPRM function with primary input signals 1=
{X n -l,X n -2,'"
,xo} and
corresponding signal probabilities {Pr,Xn-llPr,x n-2"" ,Pr,xo}, based on the above discussion, the
procedure for the power estimation of an FPRM function is as follows.
1. Given an n-variable FPRM function and its primary input signal probability set, construct
a tree of two input AND gates and XOR gates;
2. Compute the output probability for each gate using equations 7.2 and 8.5;
3. Compute the node switching activity using equations 7.5 or 7.6;

4. Compute the total power dissipation using equation 7.1.
However, for a given FPRM function with given input signal probability set, the decomposition can have a significant impact on the amount of power dissipation.
8.3.2
Area and power estimation
For an n variable function, there are 2n polarities and correspondingly there are 2n FPRM forms,
which results in different number of 'if-terms. If implemented as a tree, then the different number
of 'if-terms will have different tree sizes and result in different power dissipation. Hence, area
and power minimization of FPRM functions can be implemented by searching the best polarity
among 2n polarities. This is guaranteed by the following Lemma.
Lemma 8.1. For any n-variable FPRM function, there are 2n different forms. Each form can
be identified by corresponding polarity. Among 2n different forms, there exists at least one form,
which has the minimal number of 'if - terms. For convenience, it is called minimal form.
An n-variable FPRM function has 2ft different forms. Different forms have different function
sizes or different 'if - terms. Then, the number of 'if - terms is used to measure the area of an
128
FPRMform.
For power estimation, it is an input relevant pattern problem. Two input AND gates and
XOR gates are used to build a tree to implement an FPRM function. Then, for a given input
signal probability set, Mini-Huffman algorithm in Definition 7.6 & Algorithm 7.1 are used to
combine the input signal probabilities to calculate the total SA.
Lemma 8.1 also shows that for any n-variable FPRM function, there is one or multiple
minimal forms. If the function has one minimal form, the task of area minimization is to find a
polarity, which has the minimal form. If the function has multiple minimal forms, then choose
any of them because there is no difference among them for area minimization. However, for
power dissipation, different minimal forms may have different power dissipation for a specific
input signal pattern. Furthermore, the minimal form does not always have the minimal power.
The power dissipation under the minimal form may be higher than that under other forms. This
can be shown by the following three variable function example.
Example 8.2. A three-variable function,
bilities (PX2,PXllP XO)
1(X2, Xl, xo),
with corresponding input signal proba-
= (0.10, 0.90, 0.90) is shown as two-level FPRM format in equation 8.11.
1(X2, Xl, xo)
= Xo EEl Xl EEl XIXO EEl X2XO EEl X2 X l
(8.11)
The function is given under polarity O. It can be found that the minimal form is under
polarity 4 as shown in equation
1(X2, Xl, Xo)
= XIXO EEl X2XO EEl X2 X l
(8.12)
Also the function form under polarity 5 can be obtained as in equation
1(X2, Xl, Xo)
= Xl EEl XIXO EEl X2 EEl X2XO EEl X2 X l
(8.13)
129
X
2
Xl
Xo
Xo
Xl
x2
Xo
x2
Xl
Xl
x2
xl -T-\
x2
Xo
'" 0.18
(b)
(a)
Figure 8.2: Power dissipation with the function forms (a) The minimal form under polarity 4;
(b) The function form under polarity 5
Equations 8.12 & 8.13 can be implemented using two-input AND gates and XOR gates as
shown in Figure 8.2. The switching activity of Fig. 8.2 (a) is 6.714 while it is 2.323 for Fig.
8.2 (b). For the form under polarity 5, the power reduction is 65.4% compared to that of the
minimal form though the penalty of area is two
7r -
terms. To find a good compromise of area
and power dissipation, a cost function is proposed to guide the optimization as in equation 8.14.
Cost
= a * Area + (1 -
a)
* SA
Here, SA stands for the Switching Activity and a is the weight of area and SA and 0
(8.14)
< a < 1.
In order to find the best polarity, which gets area and switching activity minimized, among 2n
polarities, a polarity conversion algorithm is needed.
8.3.3
Polarity conversion
For area minimization, various methods have been developed. They can be classified into two
categories [92]:
1. Gray code: Sequentially search 2n polarities to find the best one which has the minimum
number of products. Memory requirement is O(2n) while the computation time is O(4n).
130
2. Extended truth vector: Using an extended truth vector and a weight vector, obtain the
number of products for 2n different expansions simultaneously. Both memory requirement and
the computation time are O(3 n ).
Any n-variable FPRM function can be uniquely expressed by its on set coefficient. Recently,
an exact method was reported in [114] to find the best polarity targeting area minimization based
on on-set coefficients by gray code, which is suitable for large functions. Memory requirement
and the computation time are O(M) and O(2n M) respectively, where M is the average number
of on-set coefficients. This method is used here to implement function polarity conversion.
For convenience, some definitions are given as follows.
Definition 8.2. Two integers can be expressed by binary n-tuples, i = {i n - 1 i n -2'" io}, j =
{jn-dn-2'" jo}. If ik
jk for all k, 0
n - 1, then i covers j or j is covered by i.
Algorithm 8.1. Wang's algorithm [116]
Given an on-set Reed-Muller coefficient set Rp for an n-variable Boolean function with polarity
p. A coefficient set Rp with any other polarity q can be achieved through the following operations
on Rp itself, where p /\ q = r (/\ is bitwise operator).
(1) For any coefficient i in the set Rp, if i does not cover r, then i is an element of Rp. Leave
i in the set. If i covers r, search the set Rp for the coefficient (i - r). If there is such a coefficient,
then delete coefficient (i - r) from the set Rp. Otherwise, if there is not such a coefficient, then
add coefficient (i - r) to the set Rp.
(2) The new set obtained in step (1) is the on-set Reed-Muller coefficient set with polarity p.
Using the above polarity conversion algorithm, the optimization implementation of FPRM
functions can be described as follows:
Given an n-input PPRM function with primary input signals 1=
corresponding signal probabilities {Pr,xn-l' Pr,x n-2'
input XOR gates are used to construct a tree T
...
{Xn-l, X n -2,'"
,xo} and
,Pr,xo}' two input AND gates and two
(V, E) with Pr,xn-ll Pr,x n-2'
... ,
and Pr,xo as
its leaves. Search the best polarity such that both area and equations 8.15 or 8.16 are minimized
131
for dynamic logic and static logic respectively.

For dynamic logic,
SA = L2Pr,i
(8.15)
iEV
For static logic,
SA =
L 2Pr,i(1 -
(8.16)
Pr,i)
iEV
Where SA is the total switching activity and Pr,i is the signal probability of an internal node.
8.4
Proposed algorithms
Based on the above discussion, a new algorithms for minimizing area and power dissipation of
single output FPRM functions is proposed as follows.
Algorithm 8.2. Power minimized algorithm

====================================================================================================
Given an on-set Reed-Muller coefficient set Rp of a single output FPRM function with the input signal probability set under polarity p, let one real variable and one integer variable, SA Cost,
AreaCost, represent the total switching activity and area cost (measured by the number of 1r-terms)
under polarity p, and let one real variable and two integer variables, BestSACost, BestAreaCost
and BestPolarity, represent the lowest total switching activity, area cost and the polarity corresponding to the lowest total switching activity, respectively. Then, for any i, 0
< i :::; 2n
1,
carry out the following steps,'

(1) Generate a polarity qi in gray code order, that is adjacent with
qi-l.
Let r
= qi /\ qi-l'
(2) Pass R qi _ 1 and r to Algorithm 8.1 to get the new on-set coefficient set R qi . Obtain
AreaCost from the set R q;.
(3) If BestAreaCost
< AreaCost, let qi = qi-l and qi =
qi
+ 1,
go to Step 1.
(4) If BestAreaCost
132
AreaCost, construct an optimal tree of two input AND gates using
Algorithm 7.6 for each 1f'-term in the FPRM function. Compute the switching activity for all
AND-trees and store it into SACost and save the primary output signal probabilities of ANDtrees into an array called ANDOutProb[}. Taking the signal probabilities in ANDOutProb[} as
the primary input signal probabilities, construct an optimal tree of two input XOR gates using
Algorithm 7.1. Compute the switching activity of the XOR-tree and add it to SACost. If SA Cost
is less than BestSACost, then BestSACost, BestAreaCost and BestPolarity are substituted by
SACost, AreaCost and qi respectively. Otherwise, go to Step 1.
(5) Output (BestPolarity /\p) that is the best polarity with BestSACost and BestAreaCost.
The above algorithm supposes that the area has the higher weight than the power dissipation.
If the function has many area minimal representations, the algorithm can find the best polarity,
which has the lowest power dissipation.

Example 8.3. A three-variable function,
with corresponding input signal proba-
f(x2, Xl, xo),
bilities (0.10, 0.45, 0.90) is shown as two-level FPRM format in equation 8.17.
f(X2, Xl, Xo)
Xo
EB
X1XO
EB
(8.17)
X2X1XO
Obviously, the function representation is under polarity O. The on-set Reed-Muller coefficient
set Ro is (1,3, 7). First, construct an optimal tree of two input AND gates using Algorithm 7.6
for each 1f'-term in the FPRM function and compute the switching activity for all AND-trees and
store it into SACost and save the primary out signal probabilities of AND-trees into an array
called ANDOutProbD. For this example, 1f'1
Xo,
1f'3
X1XO,
1f'7
X2X1XO
AND trees are shown in Fig. 8.3 (a). The switching activity is SACost
and corresponding
0.9810. The primary
out signal probabilities of AND-trees are ANDOutProb[O] = 0.9, ANDOutProb[l] = 0.405 and
ANDOutProb[2]
0.0405.
Then, take the signal probabilities in ANDOutProbD as the primary input signal probabilities,
construct an optimal tree of two input XOR gates using Algorithm 7.1 and compute the switching
133
activity of the XOR-tree and add it to SACost. The XOR tree is shown in Fig. 8.3 (b) and
add its switching activity to SACost. Obtain SACost
= 2.9460. Let BestSACost be 2.9460,
BestAreaCost be 3 and polarity 0 be BestPolarity.

Xo 0.9
Xl 0.45
Xo 0.9
0.9
=0-
X2
0.1
xl
0.45
0.405
0.0405
0.405
Xo 0.9
-----~
0.9
(a)
(b)
Figure 8.3: Decomposition and AND- trees and XOR tree (a) AND trees; (b) XOR tree
Generate polarity 1 in gray code order, 001. Let r
r
= 001 /\ 000 = 1. Pass Ro = (1,3,7) and
= 1 to Algorithm 8.1 to get the new on-set coefficient set
Rl
= (0, 1, 2, 3, 6, 7). Because
AreaCost is 6, which is greater than BestAreaCost, go to Step 1 and generate gray code for next
polarity 2. Repeat this procedure till polarity 7. Finally, find that the BestPolarity is polarity
4, BestSACost is 0.2362 and BestAreaCost is 2.
Algorithm 8.3. Power and area optimisation algorithm

====================================================================================================
Given an on-set Reed-Muller coefficient set Rp of a single output FPRM function with the
input signal probability set under polarity p, let two real variables and one integer variable, Cost,
SACost, AreaCost, represent the cost, total switching activity and area cost (measured by the number of 1r-terms) under polarity p, and let two real variables and two integer variables, BestCost,
BestSACost, BestAreaCost and BestPolarity, represent the best cost, the lowest total switching
activity, area cost and the polarity corresponding to the lowest total switching activity, respectively. Then, for any i, 0
<i
(1) Generate a polarity
qi
2n
1, carry out the following steps:
in gray code order, that is adjacent with
qi-l'
Let r
= qi /\ qi-l'
134
(2) Pass R qi _ 1 and r to Algorithm 8.1 to get the new on-set coefficient set R qi . Obtain
AreaCost from the set R qi .
(3) Construct an optimal tree of two input AND gates using Algorithm 7.6 for each 1r-term in
the FPRM function. Compute the switching activity for all AND-trees and store it into SACost
and save the primary output signal probabilities of AND-trees into an array called ANDOutProbf].
Taking the signal probabilities in ANDOutProbf} as the primary input signal probabilities, construct an optimal tree of two input XOR gates using Algorithm 7.1.
(4) Compute the switching activity of the XOR-tree and add it to SA Cost and take the number
of 1r-terms as area. Use equation 8.14 to calculate the cost. If Cost is less than BestCost, then
BestCost, BestAreaCost and BestPolarity are substituted by Cost, AreaCost and qi respectively.
Otherwise, go to Step 1.
(5) Output (BestPolarity /\p) that is the best polarity with BestSACost and BestAreaCost.
8.5
The proposed algorithm has been implemented in C and compiled by the GNU C compiler. The
time complexity for mini-Huffman is O(nlogn) [131] while that for polarity conversion is O(M2n)
[114]. Hence, the proposed algorithm has the time complexity of O(M2nnlogn) where n is the
number of variables and M is the average number of on-set coefficients. The algorithm is tested
on a personal computer with PIlI 550 CPU and 64M RAM under Linux operating system. Ten
single output circuits with input size from 8 to 21 from MCNC benchmarks are tested.
The algorithm proposed in [115] is used to convert the test circuits directly from two-level
PLA format to two-level FPRM format with polarity O. Then, the proposed algorithm will
work with the two-level FPRM format circuits. Given a set of input signal probabilities, the
best polarity is searched and its corresponding switching activity and area are found, where
area is measured by the number of 1r-terms. Twenty-one input signal probabilities are randomly
generated, which are 0.95,0.08,0.78,0.25,0.18, 0.77, 0.75, 0.71, 0.47, 0.07, 0.23, 0.73, 0.28, 0.73,
0.28, 0.73, 0.44, 0.81, 0.20, 0.38, 0.71, respectively. For different input sizes, the input signal
135
probabilities are chosen from the left to the right in the above list. For example, for three inputs
(X2'
Xl,
xo), the signal probabilities are 0.78, 0.08, 0.95 while for four inputs (X3, X2,
Xl,
xo),
they are 0.25, 0.78, 0.08, 0.95, and so on.

Two sets of experiments are conducted. The first set is based on Algorithm 8.2. Ten circuits
from MCNC benchmark are tested. Table 8.1 shows the experimental results [127]. To see the
efficiency of the algorithm, the switching activity and area under the polarity 0 are also listed
in the same table for comparison. In Table 8.1, Column 1 shows the circuit name while Column
2 shows the number of inputs (IN#) and the number of product terms (p#) in PLA format;
Columns 3 and 4 show the switching activity ( SAo) and area (Areao the number of
1r-
terms
in two-level RM format) under polarity 0 while Columns 5, 6, 7 and 8 show the best polarity,
switching activity (SA b), area (Areab) and CPU time for the best polarity using the proposal
algorithm, respectively; Columns 9 and 10 show the improvement percentage of the switching
activity and area under the best polarity compared to those under the best polarity given in
[115]. The improvement percentage of switching activities is defined as follows:
Improvement
SAo - SAb
SAo
%
(8.18)
while the one for area is similarly defined in equation 8.19.
Improvement
= Areao - Areab
'reao
(8.19)
From the results reported in Table 8.1, it is found that the switching activities can be improved
significantly for all ten circuits compared to those under polarity 0, which can be as high as 91.9%.
The average reduction of switching activities is 61.0% while the average area reduction is 32.0%
for all ten circuits. For seven cases of ten circuits, both area and switching activity are reduced.
For each circuit, 10 sets of input probabilities are randomly generated and then run on the
above algorithm. For each instance X, let Opt(X) represents the optimal solution under the best
polarity and S(X) be the solution with polarity O. The following parameter defined in equation
136
8.20 is used to measure the performance of the algorithm.
S(X) - Opt(X) o/c

S(X)
0
(8.20)
From the definition in 8.20, obviously, the higher the rate R, the worse the performance of the
algorithm. For each circuit, among 10 problems, the maximum ratios and the average improvement ratios of switching activity and area are computed, which are represented by M axSAR,
M axAreaR, AvergAreaR and AvergSAR, respectively. The results are shown in Table 8.2.
The second experiment is conducted based on Algorithm 8.3, which uses the cost function
in equation 8.14 to guide the optimization. To determine the weight a in equation 8.14, an
experiment is conducted as follows. An input signal probability set is randomly generated, in
Table 8.1: Best polarity and switching activity for low power decomposition[127]
Polarity 0
Best Polarity
Improvement (%)
Name
IN#/p#
Area
Areao polarity SA b Areab time(s) SA
SAo
14
0.21
16.84
150
2.10
87.5
75.4
Newwill
57
8/8
71.4
2.18
0.14
8.20
21
160
6
73.4
Newtag
8/8
20.54
557
29.42
84.0
-109.4
sym10
128.76
266
966
10/837
19.7
3.14
6
2.52
5
xor5
0
0
5
5/5
210
450
11.58
173
1.94
91.5
17.6
9sym
136.36
9/87
184
255
23.12
100
3.82
79.6
45.6
life
113.26
9/512
41
39323
9.82
19
29.46
76.7
53.6
t481
46.96
16/481
49152
13.74
64
16.70
41.2
20.0
ryy6
23.38
80
16/122
819200 17.22
81
41.30
87.6
50.3
cm150a
139.30
163
21/17
81
40.14
48.1
mux
77.40
786432 40.06
81
0
21/36
22.0
Average improvement (%)
73.2
Name
Newill
Newtag
Sym10
xor5
9sym
life
t481
Table 8.2: Reduction ratios of area and switching- activity

MaxSWR(%) MaxAreaR(%) AvergSWR(% ) AvergAreaR(% )
I
75.4
79.2
74.0
89.7
93.2
88.0
68.0
72.2
J
93.6
-112.0
74.2
-98.4
19.7
0
8.6
0
I
93.2
17.6
83.2
17.6
I
95.4
45.7
34.8
79.7
89.1
68.3
61.0
78.9
I
!
137
which the elements are 0.15, 0.97, 0.48, 0.10, 0.24, 0.16, 0.60, 0.96, 0.78, 0.56, 0.26, 0.74, 0.53,
0.87, 0.37, 0.39 respectively for a sixteen input variable set. Power versus alpha and area versus
alpha are conducted. The results are shown in Figs. 8.4 (a) and (b). Fig. 8.4 (a) depicts the
normalized power of four benchmarks versus alpha, i.e. the power when a
= 1 is divided by the
power with specific a. Fig. 8.4 (b) depicts the normalized area versus alpha, i.e. the area when
a = 0 is divided by the area with specific a. For circuits life & 9sym, power and area do not vary
with a while for circuits t481 & sym10, power and area varies significant with a. From Figs.
8.4 (a) & (b), it can be seen that setting a
= 0.4 produces the optimization implementations for
area and power.

Table 8.3 shows the results of ten MCNC benchmark circuits. The results of the area minimized algorithm in [116] are also listed in the table for comparison. It can be seen that two
circuits have significant power improvement without any area penalty while one circuit has power
improvement with small area trade. For the circuits 9sym and cm150a, the proposed algorithm
finds the best polarities, which have the minimal area and power dissipation. However, for the
circuit ryy6, it means that the polarity which has the minimal area is not the best polarity for
power minimization.
Name
Inputs
9sym
ryy6
mux
cm150a
9
16
21
21
Name
9sym
ryy6
cm150a
Table 8.3: Power dissipation and area comparison

Proposed in [128]
Results in [116]
SA
time (s)
P
area
SA
P
area
0.21
173
31.1
15
173 154.92
195
57344
64
19.9
16.83
49152
38.76
72
40.14
81
81
19.86
19.86
0
0
82
25.28
40.35
1
82
47.90 32768
Reduction
Area (%) SA (%)
0
79.9
-12.5
48.7
0
0
47.2
0
Table 8.4: Reduction ratios of area and SA

MaxSAR (%) MaxAreaR (%) AverageSAR (%) AverageAreaR (%)
60.0
0
79.7
0
-12.5
11.7
-3.1
48.7
-0.1
-1.2
29.9
78.0
Once again, for each circuit, 10 sets of input probabilities are randomly generated and then
run on the proposed algorithm and area efficient algorithm in [116] to see the efficiency of the cost
function in equation 8.14. Equation 8.20 is used to measure the performance of the proposed
138
algorithm. The results are shown in Table 8.4. It shows that the proposed algorithm has a
significant power dissipation improvement with smaller area penalty compared with the results
from the area minimized algorithm.
1.2
1
0.8
=
Q
iN
_.t-. life
_ _ ..JJt
-E~ 0.6
.-
6 .... 0.4
z
....
...... sym10
0.2 - -..t.. .........t..- ....

0
O 0.2 0.4
~
Q
C-
___ t481
I
I
:t;
0.6
0.8
9sym
alpha
-.a
(1;1
....
1
\
0.8
.......
'--.
.
:-- ---... ....-..
~ 0.6
..........................
---- ---- ----
0.4
Q 0.2
(1;1
Q)
- .... -life
- -- - t481
. - - i....
sym10
t:
9sym
.a
0.2
0.4
0.6
0.8
alpha
-
...... -
Figure 8.4: Power and area versus alpha (a) Power versus alpha; (b) Area versus alpha
8.6
139
Summary
A frame to estimate power dissipation of FPRM functions has been presented. Using the polarity conversion under the assumption that both signals and their complements are available, a
solution to minimize power dissipation for single output FPRM functions is proposed. The time
complexity of the proposed algorithm is O( M2nnlogn). Although the dynamic logic is taken to
explain how the algorithm works, it should be pointed out that the algorithm can be applied to
static logic as long as the calculation formula of switching activity in equation 7.5 is substituted
by the one for static logic in equation 7.6.
Chapter 9
Conclusions and Future work

This project aims to develop low power design techniques for digital logic circuits. The main
contributions in this thesis can be summarized as follows.
9.1
Low power state assignment
Using the state assignment to reduce area or power of FSMs is not new. This can be classified
as a single object optimization problem. However, using state assignment to optimize power and
area is relatively new, which is a multiple object optimization problem and much more difficult
than single object optimization. For the optimization problem, the optimization quality of a
problem mainly depends on the algorithm used and the cost function developed. An algorithm
which has large search space is essential to solve large size problems. Considering this, a genetic
algorithm is employed to do this job in Chapter 3. Corresponding operators are developed for
this specific application. Switching activity is used to estimate power dissipation of an FSM,
which includes the power dissipation both from the register section and combinational section.
Based on the number of cubes and switching activity, two cost functions are developed, which
can optimize FSMs targeting area or power dissipation or both. For a specific state assignment,
the area of the combinational section is calculated by ESPRESSO minimization. The algorithm
is implemented in C and the efficiency is tested on MCNC benchmarks. The results show that
the proposed algorithm has significant improvement in area and power dissipation compared to
140
CHAPTER 9. CONCLUSIONS AND FUTURE WORK
141
those from the state-of-the-art tools.
9.2
Low power flip-flop designs
Pulse sampling flip-flops usually suffer from a long clock chain, which is used to generate narrow
pulse, and the long clock chain consumes significant power dissipation compared to the flip-flop
cell. In Chapter 4, a clock-gating technique is used to reduce the redundant transition in a clock
chain. A novel low power Single Edge-Triggered Flip-Flop with Clock-Gating (CG-SETFF) is
proposed. The proposed flip-flop can detect the idle of the input signal to gate the clock chain
so that the clock power can be saved. From the simulation result, the power savings can be as
high as 86% compared with the published design when the input is in idle conditions.
Clock signal is the most frequent transition signal in a flip-flop except glitches. Traditional
flip-flops are single edge-triggered flip-flops, which are sensitive to the rising or falling edge
of the clock.
Therefore, half of the clock's transitions have nothing to do with the circuit
and became redundant. One solution to reduce the redundant transition is to develop double
edge-triggered flip-flops. In Chapter 5, two versions of differential CMOS double edge-triggered
flip-flops are proposed: Clock Chain Based Double Edge-Triggered Flip-Flop (CCB-DETFF)
and Pass Transistor Based Differential CMOS Double Edge-Triggered Flip-Flop (PTB-DETFF).
CCB-DETFF has the lower power dissipation than CG-SETFF when input signal probability is
greater than 0.38. PTB-DETFF only uses one clocked nMOS transistor. Therefore, it has some
advantages to use in the reduced clock swing and voltage scalability.
Reducing the number of signal lines can reduce the total capacitor and hence power dissipation
can be reduced. One solution to do that is to use multiple valued signals. In Chapter 6, multiple
valued flip-flop is explored and a novel quaternary flip-flop is proposed. The simulated result
shows that the proposed flip-flop has a correct function and has lower power dissipation compared
to other multiple flip-flop
9.3
142
Power optimization of FPRM functions
Low power techniques of AND/OR operator based domain have been extensively investigated.
However, for AND/XOR operator based domain, the research is still in its infancy.
Before technology mapping, a multiple input XOR gate needs to be decomposed into a tree
of two input gates. In Chapter 7, low power XOR gate decomposition is discussed. It is shown
that previous algorithms for low power XOR gate are not optimal. Based on polarity searching
and some given properties, a novel algorithm is presented and implemented in C. The results
show that improved power dissipation is obtained.
In Chapter 8, the above idea is extended to Fixed Polarity Reed-Muller (FPRM) functions.
A power estimation frame for FPRM functions is proposed. A cost function, which is linearly
combined by the number of 1r-terms and power dissipation, is proposed to optimize area and
power dissipation. Based on polarity conversion, two algorithms are presented. One is to optimize
power dissipation and the other is to optimize area and power. The experimental results for
MCNC benchmark circuits show that an improvement is obtained compared to the unoptimized
circuits or published results.
9.4
Future work
The above work can further be generalized and improved along the following lines.
Low power state assignment program described in Chapter 3 can be incorporated into
other logic minimizers, such as SIS, to improve their performance. The state assignment
technique can be extended to solve the multilevel FSMs in which their combinational
sections are multilevel.
It is also possible to apply this technique to low power FSM
decomposition[29] .
The idea of clock-gating technique proposed in Chapter 4 can be applied to low power
sequential circuit design[120] and low power finite state machine partitioning[18].
Multiple-valued clock based flip-flops can be further explored to design flip-flops with simpler structure. Moreover, the design methodology of low power sequential systems based
143
on this kind of flip-flops needs to be studied .

Low power XOR gate decomposition algorithm in Chapter 7 can be modified to take the
delay into account so that the target circuit is optimized in both power dissipation and
delay. If a proper delay model is applied, gliching power can be taken into account. Then,
using this logic synthesis technique, the "real" power savings will be more accurate .
Low power FPRM function algorithm in Chapter 8 is for single output functions, which
can be extended to solve multiple output FPRM functions. Further, if some algebraic
abstraction and reduction rules are applied, then the above algorithm will be more efficient
to optimize area and power.
Publications
The following are papers published, accepted and submitted while at Napier University
Y Xia and A E A Almaini, Differential CMOS edge-triggered flip-flop with clock-gating,

Electronics Letters, Vol. 38, No.1, pp.9-11, 2002
Y Xia and A E A Almaini, Genetic algorithm based state assignment for power and area
optimization, IEE Proceedings - Computers and Digital Techniques, Vol. 149, No.4, pp.
128-134, 2002
Y Xia and A E A Almaini, Best polarity for low power XOR gate decomposition, Euromi-
cro Symposium on DSD'2002 Digital System Design, Proceedings of DSD'2002, pp. 53-59
, Germany, Sept., 2002
P. Wang, X. Wu and Y Xia, Design of ternary Schmitt triggers based on its sequential
characteristic, IEEE International Symposium on Multiple-valued logic, IEEE Proc. of
ISMVL, pp. 156-160, Boston, USA, May 2002
Y Xia, A E A Almaini and Xunwei Wu, Genetic algorithm based finite state machine
optimization, Journal of Electronics (Accepted)
Y Xia and A E A Almaini, Power minimisation of FPRM functions based on polarity
conversion, Journal of Computer Science Technology (Accepted)

Y Xia and A E A Almaini, Area and power optimisation of FPRM function based circuits,
IEEE International Symposium on Circuits and Systems, Bangkok, May, 2003 (accepted)
Y Xia and A E A Almaini, A novel multiple valued flip-flop employing multiple valued
clock, (submitted to International Journal of Electronics)
144
References and Bibliography
[1] Afghahi M. and Yuan J., Double edge-triggered D-flip-flops for high-speed CMOS circuits,
IEEE Journal of Solid-state Circutis, Vol. 26, No.8, pp.1168-1170, 1991
[2] Alidian M., Monteiro J., Devadas S., Ghosh A., and Papaefthymiou M., Precomputationbased sequential logic optimization for low power, IEEE Trans. on VLSI Systems, Vol. 2,
No.4, pp. 426-436, 1994
[3] Almaini A. E. A. Electronic logic systems, Third Edition, Prentice Hall, UK, 1994.
[4] Almaini A. E. A. and McKenzie L., Tabular techniques for generating Kronecker expansion,
lEE Proc. -Comput. Digit. Tech., Vol. 143, No.4, pp. 205-212, 1996
[5] Almaini A., Miller J., Thomson P. and Billina S., State assignment of finite state machines
using a genetic algorithm, IEE Proc.-Comput. Digit. Tech., Vol. 142, No.4, pp. 279-286,
1995
[6] Almaini A. E. A. and Ping S., Algorithm for Reed-Muller expansions of Boolean functions and optimization of fixed polarities, The fourth IEEE International Conference on
Electronics, Circuits and Systems, Cairo, pp 148-153, Dec., 1997
[7] Almaini A. E. A., Thomson P. and Hanson D., Tabular techniques for Reed-Muller logic,
Int. J. Electronics, Vol.70, No.1, pp. 23-24, 1991
[8] Almaini A. E. A. and Zhuang N., Using genetic algorithms for the variable ordering of
Reed-Muller binary decision diagrams, Microelectronic Journal, Vol. 24, No.4, pp. 471480, 1995
145
BIBLIOGRAPHY
146
[9] Almaini A. E. A., Zhuang N. and Bourset F., Minimisation of multiouput Binary Decision
Diagrams using hybrid Genetic Algorithms, IEE Electronic Letters, Vol. 31, No. 20, pp.
1722-1723, 1995
[10] Ashar P., Devada S., and Newton A., Sequential logic synthesis, Kluwer, Boston, 1991
[11] Athas W., Svensson L., Koller, Tzartzanis E. and Chou E., A framework for practical lowpower digital CMOS systems using adiabatic-switching principles, International Workshop
on Low Power Design, pp. 189-194, April, 1994
[12] Bacchetta P., Daldoss L., Sciuto D. and Silvano C., Low-power state assignment techniques
for finite state machines, IEEE International Symposium on Circuits and Systems, Gnenva,
pp. 28-31, 2000
[13] Bahar R., Frohm E., Gaona C., Hachtel G., Macii E., Pardo A. and Somenzi F., Algebraic
decision diagrams and their application, Proceedings of the IEEE/ ACM International Conference on Computer-Aided Design, Santa Clara, USA, pp. 188-191,1993
[14] Benini L. and Micheli G., Dynamic power management design techniques and CAD tools,
Kluwer Academic Publishers, USA, 1998
[15] Benini L. and Micheli G., State assignment for low-power dissipation, IEEE Journal of
solid State Circuits, Vol.30, No.3, pp. 258-268, 1995
[16] Benini L. and Micheli G., Automatic synthesis of low-power gated-dock finite-state machines, IEEE Trans. on CAD Design of Integrated Circuits and Systems, Vol. 15, No.6, pp.
630-643, 1996
[17] Benini L. and Micheli G., Synthesis of low-power selectively-docked systems from highlevel specification, AMC Trans. on Design Automation of Electronic Systems, Vol. 5, No.3,
311-321, 2000
[18] Benini L., Micheli G. and Vermeulen F., Finite-state machine partitioning for low power,
Proceedings of IEEE International Symposium on Circuits and Systems, Monterey, CA,
pp. 5-8, May, 1998
BIBLIOGRAPHY
147
[19] Blair G., Low-power double-edge triggered flip-flop, IEE Electronics Letters, Vol. 33, No.
10, pp. 845-847, 1997
[20] Brace, K., Rudell R, Bryant R, Efficient implementation of a BDD package, Proceedings
of the Design Automation Conference, Orlando, USA, pp. 40-45, June, 1990
[21] Brown S., Francis Rand Vranesic Z. G. File-programmable gate array, Kluwer Academic
Publisher, Boston, 1992.
[22] Bryaton R, Graph-based algorithms for Boolean function manipulation, IEEE Trans on
Comput., Vol. C-35, No.8, pp.79-85, 1986
[23] Brayton R K., Hachtel, G. D., McMullen C. T. and Sangiovanni-Vincentelli A. L., Logic
minimization algorithms for VLSI synthesis, Kluwer Academic Publishers, Boston, 1984
[24] Chandrakasan A. and Brodensen R, Low power digital CMOS design, Kluwer Academic
Publishers, Boston, 1995
[25] Chandrakasan A., Sheng S. and Brodersen R, Low power techniques for portable realtime DSP applications, Proceedings of the 5th International Conference on VLSI Design,
Bangalore, India, pp. 203-208, Jan. 1992
[26] Chandrakasan A., Sheng S., and Brodersen R, Low-power CMOS digital design, IEEE
Journal of Solid-State Circuits, Vol. 27, No.4, pp. 473-484, 1994
[27] Chandrakasan A., Potkonjak M., et al., Optimizing power using transformations, IEEE
Trans. on Computer-Aided Design of Integrated Circuits and Systems, Vol. 14, No.1, pp.
12-31, 1995
[28] Chang J. and Pedram M., Energy minimization using multiple supply voltages, Proceedings
of International Symposium on Low Power Electronics and Design, Monterey, Calfornia,
pp. 157-162, USA, 1996
[29] Chow S., Ho Y. and Hwang T., Low power realization of finite state machines- A decomposition approach, ACM Trans on Design Automation of Electronic Systems, VoLl, No.3,
315-340, 1996
BIBLIOGRAPHY
148
[30] Coudert 0., Berthet C., Madre J., Verification of sequential machines using Boolean functional vectors, Proceedings of IFIP Inti. Workshop on Applied Formal Methods for Correct
VLSI Design, Leuven, Belgium, pp.111-128, Nov. 1989
[31] Current K., Multiple-valued logic memory circuit, International Journal of Electronics, Vol.
78, pp. 547-555, 1995
[32] Dasgupta A. and Ganguly S., Divide & Conquer: A strategy for synthesis of low power
finite state machines, Proceedings of IEEE International Conference on Computer Design,
Texas, USA, pp740-745, Oct. 1997
[33] Devadas S., Ma H., Newton A. and Sangiovanni-Vinventelli A., MUSTANG: State assignment of finite state machines targeting multilevel logic implementations, IEEE Trans.,
Comput.-Aided Des, Integr. Circuits Syst., Vol. 7, No.12, pp. 1290-1300,1998
[34] Dickinson A., Denker J., Adiabatic dynamic logic, IEEE Journal of Solid-State Circuits,
Vol. 30, No.3, pp. 311-315, 1995
[35] Drechsler R., Theobald M. and Becker B., Fast OFDD-based minimization of fixed polarity
Reed-Muller expressions, IEEE Trans. Computers, Vol. 45, No. 11, pp 1294-1299, 1996
[36] Du X., Hatchel G., Lin B. and Newton A., MUSE: A MUltilevel Symbolic Encoding Algorithm for State Assignment, IEEE Transactions on Computer Aided Design, Vol. 10, pp.
28-38, 1991
[37] Ellis S., Power management in notebook counters, Proceedings of the Personal Computer
Design Conference, PP. 749-754, 1991
[38] Forrest J., ODE: Output direct state machine encoding, Proceedings of European Design
Automation Conference with EURO-VHDL'95 on EURO-DAC, Brighton, England, pp.600605, Dec. 1995
[39] Forth R., Molitor P., An efficient heuristic for state encoding minimizing the BDD representations of the transition relations of finite sate machines, Proceedings of the IEEE/ ACM
BIBLIOGRAPHY
149
Asia and South Pacific Design Automation Conference, Yokohama, Japan, pp. 61-66, Jan.
2000
[40] Gago A. Escano R. and Hidalgo J., Reduce implementation of D-type DET flip-flops, IEEE
Journal of Solid-State Circuits, Vo1.28, No.3, pp.400-402, 1993
[41] Geiger M. and Muller-Wipefurth T., FSM decomposition revisited: algebraic structure
theory applied to MCNC benchmark FSMs, Proceedings of the Design Automation Conference, pp. 182-185, 1992
[42] Ghosh A., Devadas S., Keutzer K. and White J., Estimation of average switching activity in combinational and sequential circuits, ACM/IEEE Design Automation Conference,
California, USA, pp. 253-259, June 1992
[43] Hachtel G., Hermida M., Pardo A., Poncino M., and Somenzi F., Re-encoding sequential
circuits to reduce power dissipation, Proceedings of the IntI Conference on Computer-Aided
Design, San Jose, USA, pp. 70-73, Nov. 1994
[44] Hachtel G., Macii E., et al., Symbolic algorithms to calculate steady-state probabilities of
a finite state machine, Proceedings of IEEE European Design and Test Conference, Paris,
pp. 214-218, Feb. 1994
[45] Harris E., Depp S., Pence W., Kirkpatrick S., Sri-jayantha M and Troutman R., Technology
directions for portable computers, Proceedings of the IEEE, Vol. 83, No.4, pp. 636-658,
1995
[46] Hartmanis J., and Stearns H., Algebraic structure theory of sequential machines, PrenticeHall, New Jersey, 1966
[47] Holland J., Adaptation in natural and artificial system, University of Michigan Press, Ann
Arbor, MI, 1975
[48] Hong S., Pard S., Hwang S. and Kyung C., State assignment in finite state machines for
minimal switching power consumption, IEE Electronics Letters, Vol. 30, No.8, pp. 627-629,
1994
BIBLIOGRAPHY
150
[49] Hossain R., Wronaski L. and Albicki A., Low power design using double edge triggered
flip-flops, IEEE Trans. VLSI Syst., Vol. 2, No.2, pp. 261-265, 1994
[50] Hopcroft J., Ullman J., Introduction to automata theory, languages, and computation.
Addison-Wesley, New York, 1979
[51] http://www.intle.com/kits/quickrefyr.htm
[52] Huang M., Kwok R., Chan S., Simplified and accurate power -analysis method for deepsubmicron ASIC designs, lEE Proc. -Circuits Devices Syst., Vol. 147, No.3, pp. 175-182,
2000
[53] Johnson M. and Roy K., Datapath scheduling with multiple supply voltages and level
converters, ACM Trans. on Design Automation Electronic Systems, Vol. 2, No.3, pp.
227-248, 1997
[54] Kawaguchi H., Skurai T., A reduced clock-swing flip-flop (RCSFF) for 63% power reduction, Journal of Solid-State Circuits, Vol. 33, No.5, pp. 807-811, 1998
[55] Keonjian E., Micropower Electronics, Pergamon Press, London, New York, 1964
[56] Liew B., et al., Electromigration interconnect lifetime under AC and pulse DC stress,
International Reliability Physics Symp., pp. 215-219, 1989
[57] Lin B. and Newton A., Synthesis of multiple level logic from symbolic high-level description languages, Proceedings of International Conference on Very Large Scale Integration,
Federal Republic of Germany, pp. 187-196, Aug., 1989
[58] Lu S. and Ercegovac M., A Novel CMOS implementation of double-edge-triggered flip-flops,
IEEE Journal of Solid-state Circuits, Vol. 25, No.4, pp. 1008-1010, 1990
[59] Kim J., Jang Y. and Park H., CMOS sense amplifier-based flip-flop with two N-C 2 MOS
output latches, lEE Electronics Letters, Vol. 36, No.6, pp. 498-499, 2000
[60] Macii E. and Pncino M., Design techniques and tools for low-power digital systems, EuroTraining Course, Italy, Nov. 2000
BIBLIOGRAPHY
151
[61] Mangione-Simith B., Low power communication protocols: Paging and beyond, Symposium on Low Power Electronics, San Jose, USA, pp.8-11, Nov. 1995
[62] Marculescu D., Marculescr R. and Pedram M., Information theoretic measures for power
analysis, IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems,
VoLl5, No.6, pp. 599-598,1996
[63] Marculescu D., Marculescr R. and Pedram M., Theoretical bounds for switching activity
analysis in finite-state machines, IEEE Trans. on VLSI systems, Vol. 8 , No.3, pp. 335-339,
2000
[64] Mckenzie L., Logic synthesis and optimization using Reed-Muller expansions, PhD thesis,
Napier University, UK, 1995
[65] Meng T., Gordon B., et aI., Portable video-on-demand in wireless communication, Proceedings of the IEEE, Vol. 83, No.4, pp. 659-680, 1995
[66] Meindl J., Micropower Circuits, J. Wiley and Sons, New York, 1969
[67] Micheli G., Brayton R. and Sangiovanni-Vincentelli A., Optimal state assignment for finite
state machines, IEEE Trans. on CAD, Vol. CAD-4, No.3, pp.269-283, 1985
[68] Mishra S., Rofail M. and Yeo, K., Design of high performance double edge-triggered flipflops, lEE Proc. -Circuits Devices Syst., Vol. 147, No.5, pp. 283-2, 2000
[69] Molhi S. and Chatterjee P., I-V microsystem-scaling on schedule for personal communications, IEEE Circuits and Devices, pp. 13-17, 1994
[70] Monteiro J. and Devadas S., Retiming sequential circuits for low power, Proceedings of the
Int'l Conference on Computer-Aided Design, California, USA, pp. 398-402, 1993
[71] Monteiro J. and Oliveira A., Finite state machine decomposition for low power, Proceedings
of the 35th Annual Conference on Design Automation Conference, San Francisco, USA,
pp. 758-763, June 1998
BIBLIOGRAPHY
152
[72] Mosisiads Y. and Bouras 1., Differential CMOS edge-triggered flip-flop based on clock
racing, Electronics Letters, Vol 36, No. 12, pp. 1012-1013, 2000
[73] Najm F., Goel. S. and Hajj 1., Power estimation in sequential circuits, ACM/IEEE Design
Automation Conference, San Francisco, USA, pp.635-640, June 1995
[74] Narayanan U. and Liu C. L. Low power logic synthesis for XOR based circuits, International
Conference on Computer-Aided-Desigm, San Jose, USA, pp. 570-574, Nov. 1997
[75] Nemani M. and Najm N., Towards a high-level power estimation capability, IEEE Transaction on Computer-Aided Design of Integrated Circuits and Systems, Vo1.15, No.6, pp.
588-598, 1996
[76] Oldfield J., Dorf R., Field programmable gate arrays, John Wiley & Sons, Inc., New York,
1995
[77] Olson E. and Kang S., Low-power state assignment for finite state machines search, Proceedings of IWLPD'94: ACM/IEEE International Workshop on Low Power Design, Napa
Valley, pp. 63-68,1994
[78] Panda P., Dutt N., Reducing address bus transitions for low power memory mapping, IEEE
Europena Design and Test Conference, Paris, pp. 63-67, March 1996
[79] Parrilla L., Ortega J. and Lloris A., Nondeterministic AND-EXOR minimisation by using
rewrite rules and simulated annealing, lEE Proc., Comput. Digit. Tech., Vol. 146, No.1,
pp.I-8, 1999
[80] Pedram M, Wu Q and Wu X, A new design of double edge triggered flip-flops, Proceedings
of the Asia and South Pacific Design Automation Conference, Yokohama, Japan, pp. 417421, Feb. 1998
[81] Perry D., VHDL, 3rd edition, McGraw-Hill Inc., New York, 1998
[82] Prosser F., Wu X., and Chen X., 1988, Ternary CMOS flip-flops and their applications,
lEE Proceedings, Vol. 135E, pp. 256-272, 1988
BIBLIOGRAPHY
153
[83] Rabaey J. and Pedram M., Low power design, Kluwer Academic Publishers, USA, 1996
[84] Raghunathan A., Dey S. and Jha N., Wakabayashi K., Power management techniques
of control-flow intensive design, ACM/IEEE Design Automation Conference, California,
USA, pp.429-434, 1997
[85] Raghunathan A. and Kja N., Behavioral synthesis for low power, Proceedings of the International Conference on Computer Design, Cambridge, USA, pp. 318-322, Oct. 1994
[86] Rashid M., SPICE for circuits and electronics using PSpice, Prentice Hall, New Jersey,
1995
[87] Reddy S., Easily testable realization for logic functions, IEEE Transaction on Computers,
No.11, pp. 1183-1188, 1972
[88] Roy K. and Prasad S., Syclop: Synthesis of CMOS logic for low power application, Proceedings of IEEE International Conference on Computer Design, Cambridge, USA, pp.
464-467, 1992
[89] Roy K. and Prasad S., Low-power CMOS VLSI circuit design, John Wiley & Sons, Inc.,
New York, 2000
[90] Roy S., Banerjee P. and Sarrafzadeh M., Partitioning sequential circuits for low power,
Proceedings of the IEEE International Conference on VLSI Design, India, pp. 212-217,
Jan. 1998
[91] Panda R. and Najm F. Technology decomposition for low-power synthesis, IEEE Custom
Integrated Circuits Conference, Santa Clara, CA, pp. 627-630, May 1995.
[92] Sasao T. and Fujita M. Representations of discrete functions, Kluwer Academic Publishers,
USA, 1996
[93] Semiconductor Industry Association, Workshop Working Group Reports, Irving, TX, 2223, 1992
BIBLIOGRAPHY
154
[94] Sentovich E. M., Singh K. J., Lavagno L. Moon C., Murgai R., Saldanha A., Savoj J.,
Stephan P. R., Brayton R. K., and Sangiovanni-Vincentelli A., SIS: A system for sequential
circuit synthesis, Electronics Research Laboratory, Memorandum No. UCB/ERL M92/41,
UC, Berkeley, May, 1992
[95] Shahidi G., et al., O.lfl,m CMOS devices, Proceddings of the Third Great Lakes Symposium
on VLSI Technology, pp.67, 1993
[96] Singh D., Prospects for low power microprocessor design, International Workshop on Low
Power Design, Napa, Cal., p. 1, 1994
[97] Song N. and Perdowski M., New fast approach to approximate ESOP minimization for
incompletely specified multi-output functions, Proc. Reed-Muller'97 Conference, Oxford
Univ., UK, pp 67-72, 1997
[98] Stan M., Burleson W., Bus-invert coding for low-power I/O, IEEE Trans on VLSI Systems,
Vol. 3, No.1, pp.49-58, 1995
[99] Stojanovic V. and Oklobdzija V. G. , Comparative analysis of master-slave latches and flipflops for high-performance and low-power system, IEEE Journal of Solid-State Circuits,
Vol. 34, No.4, pp. 536-548, 1999
[100] Strollo A. and Caro D., Low power flip-flop with clock gating on master and slave latches,
Electronics Letters, Vol. 36, No.4, pp. 294-295, 2000
[101] Strollo A, Napoli E and Caro D, Low-power flip-flops with reliable clock gating, Microelectronics Journal, Vol. 32, pp.21-28, 2001
[102] Strollo E. Napoli E. and Cimino C., Low power double edge-triggered flip-flop using one
latch, Electronics Letters, Vol. 35, No.3, pp. 187-188, 1999
[103] Thomas D., The Verilog Hardware Description Language, Fourth Edition, Kluwer Academic Publishers, Lowell, MA, 1998
[104] Tiwari V., Ashar P. and Malik S., Technology mapping for low power, Proceeding of the
Design Automation Conference, Dallas, Texas, pp. 74-79, June 1993
BIBLIOGRAPHY
155
[105] Tran A. adn Wang J., Decomposition method for minimisation of Reed-Muller polynomials
in mixed-polarity, IEE Proc-E, Vol. 140, No.1 pp. 65-68, 1993
[106] Trivedi K., Probability and statistics with reliability, queueing, and computer science applications, Prentice-Hall, USA, 1982
[107] 'Thai C. and Marek-Dadowska M., Multilevel logic synthesis for arithmetic functions, Design
Automation Conference, Las Vegas, pp. 68-73, June 1996
[108] 'Thui C., Pedram M. and Despain A., Low power state assignment targeting two- and multilevel logic implementations, IEEE Trans. on Computer-Aided Design of Integrated Circuits
and Systems, Vol. 17, No.12, pp. 1281-1291, 1998
[109] Tyagi A., Entropic Bounds on FSM switching, IEEE Transactions on VLSI Systmes, Vol.5,
No.4, 456-464, 1997
[110] Unger S., Double-edge-triggered flip-flops, IEEE Tans. on Computers, Vol. C-30, No.6,
pp.447-451, 1981
[111] Villa T., Sangiovanni-Vincentelli A., NOVA: State assignment of finite state machines for
optimal two-level logic implementations, IEEE Trans. Computer-Aided Design of Integrated Circuits and Systems, Vol. 9, No.9, pp. 905-924, 1990
[112] Vranesic, Z., Multiple-valued logic: an introduction and overview, IEEE Transactions on
Computers, Vol. C-26, pp. 1181-1182, 1977
[113] Wang K., Wang W., Hwang T., Wu A. and Lin Y., State assignment for power and area
minimization, Proceedings of IEEE International Conference on Computer design, Cambridge, USA, pp. 250-254, Oct. 1994
[114] Wang L., Automated synthesis and optimization of multilevel logic circuits, PhD thesis,
Napier University, UK, 2000
[115] Wang L. and Almaini A. E. A., Fast conversion algorithm for very large Boolean functions,
Electronics Letters, Vol. 36, No. 16, pp. 370-1371, 2000
BIBLIOGRAPHY
156
[116] Wang L. and Almaini A. E. A., Exact minimisation of large multiple output FPRM functions, lEE Proc.-Comput. Digit. Tech., Vol. 149, No.5, pp. 203-212, 2002
[117] Wang, S. and Horng, M., State assignment of finite state machines for low power applications, Electronics Letters, Vol. 32, No. 25, pp. 2323-2324, 1996
[118] Weste N., and Eshragian K., Principles of CMOS VLSI design: a systems perspective,
Second Edition, Addison-Wesley Publishing Company, Santa Clara, 1992
[119] Woodford C., Solving linear and non-linear equations, Ellis Horwood Limited, England,
1992
[120] Wu Q., Pedram M and Wu X, Clock-gating and its application to low power design of
sequential circuits, IEEE Trans. on Circuits and Systems, Part 1, Vol.47, No.3, pp. 415420, 2000
[121] Wu X., and Chen X., Quaternary CMOS circuits based on transmission function theory,
Chinese Science (Series A), No.5, pp. 528-536, 1989
[122] Wu X., Shen J., and Chen X., CMOS multivalued flip-flops based on new presetting scheme
and transmission function theory, Proceedings of International Workshop on Spectral Technique, Beijing, pp. 74-77, 1994
[123] Wu X., Chen B. and Pedram M., Power estimation in binary CMOS circuits based on
multiple-valued logic, Journal of Multiple Valued Logic, Gordon and Breach Publishing
Group, Vol. 7, No. 3-4, pp. 195-211, 2001
[124] Xia Y. and Almaini A. E. A., Genetic algorithm based state assignment for power and area
optimisation, IEE Proc.-Comput. Digit. Tech., Vol. 149, No.4, 128-133, 2002
[125] Xia Y. and Almaini A. E. A., Differential CMOS edge-triggered flip-flop with clock-gating,
Electronics Letters, Vol. 38, No.1, pp. 9-11, 2002
[126] Xia Y. and Almaini A. E. A. , Best polarity for low power XOR gate decomposition,
Proceedings of Euromicro Symposium on Digital System Design, Dortmund, Germany, pp.
53-59, Sept. 2002
BIBLIOGRAPHY
157
[127] Xia Y., Almaini A.E.A. and Wu X., Power minimization of FPRM function on polarity
conversion, Journal of Computer Science Technology (Accepted)
[128] Xia Y., Almaini A.E.A., Area and power optimization of FPRM function based circuits,
IEEE International Symposium on Circuits and Systems, Bangkok, Thailand, May 25-28,
2003 (Accepted)
[129] Yang S., Logic synthesis and optimization benchmarks user guide, Microelectronics Center
of North Carolina, 1991
[130] Yim J. and Kyung C., Datapath layout optimization using genetic algorithm and simulated
annealing, lEE Proc.-Comput. Digit. Tech., Vol. 145, No.2, pp.135-141, 1998
[131] Zhou H. and Wong D. F., Optimal low power XOR gate decomposition, Design Automation
Conference, Bergen, Norway, pp.104-107, Aug. 2000
[132] Zhuang N. and Wu H., Novel ternary JKL flip-flops, Electronics Letters, Vol. 26, pp.
1145-1146, 1990
[133] Zukeran, C., et aI, 1985, Design of new low-power quaternary CMOS logic circuits based on
multiple ion implants, IEEE International Symposium on Multiple-Valued Logic, Kingston,
pp. 84-90, 1985
Disk Containing the Software

The attached floppy disk contains the programs developed in the previous chapters.
GA based FSM low power state assignment: Shell script file and C source files
GA based FSM power and area optimisation: Shell script file and C source files
Low power XOR gate decomposition: C source files
FPRM function power optimisation: FPRM function power minimisation; FPRM function
power and area optimisation
158

Low Power Design Techniques For Digital Logic Circuits

Uploaded by

Copyright:

Available Formats

Low Power Design Techniques For Digital Logic Circuits

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

Low Power Design Techniques For Digital Logic Circuits

Uploaded by

Copyright:

Available Formats

Low Power Design Techniques for Digital Logic Circuits

A thesis presented in partial fulfilment

To implement a complex function involving many inputs,

Motivations of low power design. . . . . . .

VLSI chip design approaches for low power

Sources of power dissipation . . . . .

Design approaches of low power VLSI systems.

Low power design flow . . . . . . . .

Low power digital design techniques

State assignment for low power Finite State Machines (FSMs)

Low power flip-flop design . . . . . . . . . . . .

Low power logic synthesis for FPRM functions

Conventional low power design techniques

Low power design approaches of FSMs . . . . . . . . .

Finite state machines and their representations

FSM power estimation

FSM power optimization .

2.2 Low power design of flip-flops ..

Power optimization of XOR gate based circuits

State assignment for area and power optimization

Finite state machines (FSMs) . . . . . . . . . . .

Terminology and parameter calculation of FSMs

3.4 State assignment using Genetic algorithm

Outline of the algorithm

Differential CMOS single edge-triggered flip-flop

Differential CMOS single edge-triggered flip-flop.

Circuit parameter optimization

Differential CMOS Single Edge-Triggered Flip-Flop with Clock-Gating (CG-SETFF) 62

Circuit simulations and power dissipation measurement.

Low power binary counters

4-bit binary counter

Binary twisted ring counters

5 Differential CMOS double edge-triggered flip-flops

Clock Chain Based Double Edge-Triggered Flip-Flop (CCB-DETFF) using single

Pass Transistor Based Differential CMOS Double Edge-Triggered Flip-Flop (PTBDETFF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..

Voltage scaling property for CCB flip-flops.

Multi-valued flip-flop approach

Traditional Binary CP (BCP) controlled multiple valued master slave flip-flops in

Multiple valued CP (MVCP) controlled multiple valued flip-flops in parallel

7 Low power XOR gate decomposition

Definitions and terminology

Proposed low power XOR gate decomposition

Power and area optimization of FPRM functions

Definition and terminology.

Polarity based area and power minimization of FPRM functions .

Area and power estimation

Conclusions and Future work

Low power state assignment .

Low power flip-flop designs

Power optimization of FPRM functions.

References and Bibliography

Disk Containing the Software

MCNC Microelectronics Center North Carolina

Market of portable applications is growing very rapidly .

Battery capacity during the past 30 years

1.4 Charging and discharging for an inverter

Low power design flow

FSM structure model .

Monte-Carlo-based technique flow chart

~~- L_ _ _ _ _ _~_ _ _ _ _ _ _ _L -_ _ _ _ _ _~_ _ _ _ _ _~