Optical hybrid package with an 8-channel 18GT/s CMOS transceiver
for chip-to-chip optical interconnect
E. Mohammed*a, J. Liaoa, A. Kerna, D. Luc, H. Braunischc, T. Thomasb, S. Hyvonena, S. Palermoa
and I. A. Younga
a
Intel Corporation 2513 NW 229th St., Portland, OR 97124
Intel Corporation 2111 NE 25th Av., Portland, OR 97124
c
Intel Corporation 5000 W Chandler Blvd., Chandler, AZ 85226
b
ABSTRACT
We describe the design and development of a high-speed 8-channel hybrid integrated optical transceiver package with
Clock and Data Recovery (CDR) circuits. The package concept has been developed to be compatible with
microprocessor package technology and at the same time allow the integration of low cost, high-performance optical
components. A 90nm CMOS optical transceiver chip, 850nm 10Gb/s GaAs based vertical cavity surface emitting laser
(VCSEL) array and PIN photodiode array are flip-chip mounted on a standard microprocessor Land Grid Array (LGA)
package substrate. The CMOS drivers and receivers on the transceiver chip and the optical components (VCSEL and
Photodiode arrays) are electrically coupled using a short transmission line routed on the top surface of the package.
VCSEL and photodiode arrays are optically coupled to on-package integrated polymer waveguide arrays with metallized
45o mirrors. The waveguides, which are terminated with multi-terminal (MT) fiber optic connectors, couple out/in highspeed optical signals to/from the chip. The CMOS transceiver chip fully integrates all analog optical circuits such as
VCSEL drivers, transimpedance amplifiers and clock and data recovery (CDR) retiming circuit with a low jitter LCPLL. Digital circuits for pseudorandom bit-pattern sequence generators (PRBS) and bit-error rate test (BERT) are fully
integrated. 20Gb/s electrical and 18Gb/s optical eye diagrams for the transmitter were measured out of the package. A
fully packaged transmitter and receiver including clock data recovery at 10Gb/s have also been measured.
Keywords: Optical I/O, CMOS, CDR, VCSEL arrays, photodetector arrays, polymer waveguides, hybrid packaging,
flip-chip
1. INTRODUCTION
The recent transition to multi-core or many core architectures in the computer industry combined with the steady
increase of central processing unit (CPU) clock frequency brings new bandwidth demands to package-to-package (chipto-chip) electrical I/O between CPU and chipset on computer motherboards and among multiple CPUs in blade servers1.
The era of terascale computing is quickly approaching1. Traditional practices to meeting these demands require
increasing the I/O bus width and individual I/O data transmission rate. However, unless alternative ways are found, this
method of increasing I/O bandwidth eventually comes at the cost of compromising the integrity of the signal and
increasing power consumption. This section focuses on defining the underlying problems both in the personal computers
(PC) and server space using existing I/O architectures but being mindful of the potential future architectural changes in
the CPU, I/O or package. These architectural changes may shift or delay such problems in electrical I/O but optics has
the potential to provide the ultimate solution because of its good attributes of low signal propagation loss in
fibers/waveguides. Figure 1. shows the system I/O architecture in a typical modern personal computer (PC). It shows the
CPU and chipset, consisting of memory control hub (MCH) and I/O control hub (ICH) all connected to each other by
interconnect buses. The key potential future bandwidth bottleneck points, and therefore candidates for optics, are the
front side bus (FSB) that connects the CPU and MCH and the PCI Express bus that connects the graphic accelerator and
the MCH. Lower bandwidth PCI express buses are also used in the system architecture below for external I/O
communication such as Gigabit Ethernet.
*edris.m.mohammed@intel.com; phone (503) 613-4298; fax (971) 214-7807; Intel Corporation
Photonics Packaging, Integration, and Interconnects VIII, edited by Alexei L. Glebov, Ray T. Chen
Proc. of SPIE Vol. 6899, 68990Z, (2008) · 0277-786X/08/$18 · doi: 10.1117/12.764001
Proc. of SPIE Vol. 6899 68990Z-1
Inter
Pentium®
Processor
FSB
Figure 1. The current Intel system I/O architecture for x86 PC.
Figure 2. is a schematic showing the corresponding physical interconnection between the CPU and the MCH through an
individual FSB electrical I/O. As illustrated in the figure the transmitted signal from the CPU reaches the MCH after
traversing approximately 18 centimeters on the printed circuit board (PCB) and passing through several discontinuities
on both sides of the package. In this process the received signal at the MCH will suffer from frequency dependent losses,
reflections and cross-talk that will close the data eye and increases its bit-error-rate unless the interconnect is
compensated.
Cpu
MCH
PACKAGE
CPU
MCH
Printed Circuit Board
Figure 2. A schematic of interconnection between CPU and MCH through the front side bus (System Bus).
In the server space similar issues are encountered for I/O buses between CPU chips (packages) located on different
server blades. However, due to the longer length the signal travels and the additional electrical discontinuities in its path,
much more severe signal integrity issues are expected as the data transfer rate is increased. Also, optical transmission
becomes an increasingly viable solution for longer interconnect lengths. Therefore, we will examine the blade server
physical layer to understand the source of bandwidth limitation and the potential for optical interconnects to be deployed.
Figure 3. illustrates an arrangement of two server blades on a backplane connected electrically. For chip-to-chip I/O data
transmission on the backplanes, the high-speed signal driven out off-chip from the I/O bumps of CPU1 has to travel
through two multi-layer package substrates, two CPU sockets, three multilayer PCBs and two backplane connectors
before being received at CPU2 on a different blade 54 centimeters away. During this travel the signal suffers from
frequency dependent loss and encounters many discontinuities such as various pad/bump capacitances and package via
inductances as shown in Figure 3. The frequency dependant transmission losses are caused primarily by skin effect and
dielectric loss. High frequency components of the signal are greatly attenuated in comparison to lower frequencies. The
attenuation of signal on a transmission line causes intersymbol interference (ISI) because this attenuation prevents the
signal from reaching full strength within its symbol time, causing it to spread into the next bit time slot. This signal
deterioration effect will be reflected in the smearing of the eye diagram and increase in the bit-error rate at the receiver.
Proc. of SPIE Vol. 6899 68990Z-2
CPU2
CPU1
Connector
Figure 3. An example of CPU Package-to-package electrical interconnection between server blades. Colored traces on the board show
I/O signal path from CPU1 to CPU2.
Figure 4 shows simulated insertion loss data for a situation found in the above mentioned server blade for 54 cm long
electrical interconnect connecting the two CPUs. The result includes the effects of packages, pad capacitance, via
inductance and two backplane connectors in addition to the linear frequency dependent losses associated with the
transmission line traces on PCB made out of FR4 material. The insertion loss data shows a largely linear dependence on
frequency indicating the dominant factor to be dielectric attenuation. At 10GHz the simulation for this link predicts an
insertion loss of 50dB which is near the limit for electrical transmission. A dip around 10GHz is likely to be from
discontinuity due to via inductance or impedance mismatch. Generally driver pre-emphasis or various receiver
equalization techniques are used to compensate for the signal loss. Pre-emphasis operates by boosting the high frequency
energy every time there is a transition in the data, since this is where the most signal loss occurs. Receiver equalization
provides functionality in the receiver to help overcome high frequency signal losses of the transmission medium. It acts
as a high pass filter and amplifier to the data as it enters into the receiver.
0
Insertion Loss (dB)
-10
-20
-30
-40
-50
-60
-70
-80
0
5
10
Frequency (Hz)
15
9
x 10
Figure 4. Projected channel loss for electrical in electrical link in servers.
However, both methods have some associated overhead since they use additional circuitry. The success of long, medium
and short distance optical communication has proven that at high frequencies optical fibers/waveguides have very high
bandwidth compared to electrical interconnects, minimal propagation loss and negligible cross-talk. However, viable
implementation of optical interconnects in ultra-short distances in computing systems requires looking at the problem
from a new perspective. A mere replacement of physical layers may not be sufficient unless combined with a change in
some aspect of the system architecture. As most of the problems are parasitic related, the architectural change should
include very close integration of electronic and optical chips together with revolutionary packaging approaches that
Proc. of SPIE Vol. 6899 68990Z-3
emphasize miniaturization. Figure 5 is an illustrative hypothetical example for a possible approach to use optics in the
server architecture that emphasizes close integration of electronic and optoelectronic dies. In this arrangement the laser
driver and the receiver are assumed to be integrated with the CPU. They are then interfaced with the respective
optoelectronic dies (lasers or photodetectors) through short microstrip lines (3-5 mm in length) on the package as shown
below. Typical insertion losses for these microstrip lines would be ~1dB at 20GHz. Multimode fiber arrays that are
terminated with multi-terminal (MT) optical connectors physically connect the two chips to enable signal transmission
between CPU1 and CPU2. If an efficient optical coupling is achieved at every interface, there will be a potential for the
total power consumption to decrease appreciably.
cpu1
CPU2
-Microatrip
Optical
Connector
tical
Figure 5. Optical interconnection between two microprocessor units on blade servers.
The following sections describe the development of an 8-channel optical transceiver hybrid package which attempts to
demonstrate integration of CMOS and optics using standard microprocessor packaging. Section 2 discusses the
architecture and design of the chip including the key circuit blocks for optical. Section 3 will describe the package
architecture and design. Experimental results will be discussed in section 4 and a summary will be given in section 5.
2. ARCHITECTURE and DESIGN
2.1 Transceiver chip architecture
The dual transceiver chip shown in Figure 6 was designed and fabricated in Intel’s 90nm digital CMOS process
technology with seven metal layers on high-resistivity substrate2. The high resistivity reduces noise coupling through the
substrate. The total dimension of the dual chip is 5 x 10 mm2 but 1/3 of the chip is unoccupied. The chip has eight
transmitter (Tx) and receiver (Rx) channels, six of which contain key circuit blocks such as transimpedance and limiting
amplifiers (TIA, LIA), clock and data recovery (CDR), pseudorandom bit-pattern sequence (PRBS) generator and
VCSEL drivers. The two remaining Tx and Rx channels only contain TIA, LIA and VCSEL drivers and they are
intended to test these individual blocks separately using external inputs. All logic functions on the chip are implemented
using high speed differential current mode logic (CML). The differential nature of the circuits makes them immune to
power supply variations and substrate noise. The high-speed gates of CML topology have been optimized to provide
bandwidth in excess of 10GHz and they do not require any inductive peaking as used in many similar recent optical
interconnect reports3,4. Inductors are not desired because they complicate chip fabrication and require more real estate
increasing the total chip area. The center block of the chip is used for global distribution of power and bias as well as
testing scan bits and reference clock signals to all channels. An on-chip bandgap reference circuit provides bias to all the
CML circuits. Three separate power supplies were used for the core logic, the voltage control oscillator (VCO) and for
the cascode TIA. These are 1.2V, 1.4V and 1.8V, respectively. In order to minimize switching noise decoupling and
Proc. of SPIE Vol. 6899 68990Z-4
bypass capacitors were used at the chip, package and board level.
(a)
10m
Tx
Tx
Tx
Tx
Tx
Tx
Tx
Tx
5m
R
R
R
R
R
R
R
R
Rxx
Rxx
Rx
x
Rxx
Rxx
Rxx
Rxx
Rx
x
(b)
Figure 6. (a) transceiver chip layout (b) picture of actual transceiver.
Single channel architecture
Figure 7 shows the block diagram of a fully integrated single transceiver channel with all the necessary circuit blocks for
standard CDR based optical I/O transmission. This is a universal channel that is used as a transmitter (Tx) or a receiver
(Rx) depending on the configuration set by the scan chain through a computer interface. When the unit is in a
transmission mode, the key functional blocks used are the phase locked loop (PLL) with input reference clock, the
pseudorandom bit sequence (PRBS) generator and the VCSEL driver. The PLL uses an LC-VCO5 topology for its
superior low jitter performance. The VCO has four digitally controlled switch capacitors which extend its tuning range
from 7.5-11.2GHz. In a transmit mode the PLL is locked using a standard PFD loop. The clock from the PLL is then fed
to a pattern generator circuit block which produces 215-1 PRBS non-return to zero (NRZ) data signals to drive the
VCSEL driver as shown in Figure 7. The transmit block also has an option where it receives a high-speed external clock
using the input LIA/MUX circuits to drive the PRBS directly. The PRBS has the capability for half-rate sampling in
order to double the output frequency of the PRBS data stream. The receive mode allows both electrical and optical inputs
to be selected using the input MUX circuit. In optical receive mode a photocurrent from an on-package photodetector is
fed to a transimpedance amplifier (TIA) for amplification and then through a limiting amplifier (LIA) to CML level. It
then passes to a PLL based CDR unit as shown in Figure 7. The CDR has a dual loop PLL based on phase frequency
detector (PFD) and a bang-bang phase detector (BBPD). During startup it first uses the external reference clock and PFD
loop for frequency locking, as in the transmit mode described above. When the frequency difference between VCO and
reference clock is less than 500ppm, the lock signal goes high, and the PLL starts to lock to the received serial data using
a bang-bang phase detector (PD). The CDR retimes the incoming high-speed data to remove jitter. The output MUX
selects the received and retimed signal from the CDR and passes it to a 50Ω output driver for off-chip detection.
Proc. of SPIE Vol. 6899 68990Z-5
PRBS
BERT
OUTPUT
VCSEL
Driver
VCO
LOCK
PFD
BBPD
(a)
PRBS
BERT
OUTPUT
VCSEL
Driver
VCO
LOCK
LIA+TIA
PFD
BBPD
(b)
Figure 7. Single universal channel architecture (a) block diagram (b) channel layout design.
VCSEL drivers
The VCSEL devices used in this development are rated for 10Gb/s. Beyond 10Gb/s they are bandwidth limited with a
slow tail due to intrinsic and extrinsic parasitic effects such as carrier diffusion6 and device capacitance of 700fF. In
order to compensate for these effects and increase the frequency, drivers with pre-emphasis were used. The pre-emphasis
VCSEL drivers7 integrated on the transceiver chip have a dual edge pre-emphasis and generate a sub-bit-period timing
directly from the input data. A pre-emphasized current waveform is generated by summing the main modulation current
with a delayed and weighted peaking current in order to produce pre-emphasis pulses at each data transition as shown in
Figure 8. The VCSEL supply voltage is independent from the core supply to allow test flexibility. A 5-bit digital-toanalog converter (DAC) provides digital control of the output currents from 0−46mA and 0−15mA for the main and preemphasis drivers, respectively. Typical average currents provided to VCSELs range from 6−10mA which corresponds to
an average optical power of 1.5−2mW. The VCSEL driver is output terminated and connected to the VCSEL through a
50Ω microstrip transmission line that was routed on the top surface of the package. The design of the VCSEL drivers
comprehended this microstrip transmission line and the VCSEL optical and electrical models. Extracted optical and
electrical models of the VCSEL have been included to accurately simulate the effects of carrier dynamics. The detail
description of the design of this pre-emphasis VCSEL driver is available7.
Proc. of SPIE Vol. 6899 68990Z-6
MB4
MB3
MB2
MB1
MB0
5−Bit
DAC
IB1
M
IMod
Data
ISum
P
Digital
Delay Line
S1
S0
IPeak
VCSEL
Driver
IB2
PB4
PB3
PB2
PB1
PB0
5−Bit
DAC
Figure 8. Block diagram of the pre-emphasis VCSEL driver.
The transimpedance amplifier
The transimpedance amplifier integrated on each transceiver channel has symmetric feedback and has a differential
topology as shown in Figure 9. This differential topology converts the single-ended current input to a differential output
voltage to help mitigate supply noise at subsequent gain states and improve stability by allowing negative feedback from
a two-stage amplifier. The TIA was designed to have a dominant pole of 10GHz which provides a data rate in excess of
12.5Gb/s for an input parasitic capacitance of Cp=250fF. It has a feedback resistance of 314Ω and open loop gain of 3.9.
It receives a single-ended photocurrent of 200µA from the photodiode and generates a differential 2x50mVp-p output
that is fed to the limiting amplifier (LIA) which converts it to a CML level output. The LIA consists of a cascade of
buffers based on current mode logic (CML). In the first stage of the LIA, sizing the differential pair devices for optimum
gain bandwidth product minimizes input referred noise. Typically this LIA achieves an input sensitivity of 2 x 22.5mVpp. The LIA design does not require any inductive peaking since the 90nm CMOS technology2 has NMOS transistors with
an ft of at least 130GHz. A digital-to-analog (DAC) DC offset cancellation circuit is implemented at the TIA input to
remove the DC current component caused by the off-state DC bias of the VCSEL transmitter. In the packaged
transceiver the combined photodiode and pad/bump/ESD capacitances could go as high as 500fF. This increase in the
parasitic capacitance decreases the targeted data rate that could be measured in the receiver channel. At the chip level for
the same TIA an electrical eye diagram of 18Gb/s was measured previously for an input capacitance of 88fF indicating a
strong dependence on the parasitic capacitance7.
VPD
RF
PD
CF
Iph
Limitin
-A
g
V+
V-
CF
RF
IDC
CDC
F
Figure 9. Block diagram of a symmetric transimpendence amplifier connected to a photodiode
Proc. of SPIE Vol. 6899 68990Z-7
3. TRANSCEIVER PACKAGE
3.1 Package architecture and design
The package architecture has been developed to be compatible with the current microprocessor C4 organic package
technology and at the same time allow the integration of low cost, high-performance optical components. Since this
packaging approach has been discussed in detail previously, only a brief relevant description is given here8,9. Figure 10
shows the package layout design. The package substrate has a dimension of 31x31mm2 with a standard build up
consisting of laminated copper layers separated by dielectric. The electrical design of the package involves the routing of
signal and power lines as well the use of decoupling capacitors. All the high-speed electrical lines on the substrate are
routed as controlled impedance (50Ω, single-ended or 90Ω, differential) 5GHz microstrip or stripline traces. The
single−ended microstrip lines are routed on the top surface of the substrate and connect the VCSEL driver (TIA) output
bumps to VCSEL (photodiode) bumps on the package. The close proximity between them minimizes frequency
dependent loss. The length of these microstrip lines varies because of the relative size difference between the Si die and
optoelectronic chips as illustrated in Figure 10. The different layers of the substrate provide signal, power and biasing.
Bias planes were used to externally bias VCSELs and photodiodes. The decoupling capacitors on the top surface are used
to minimize switching noise on power and bias planes/lines.
Figure 10. Package substrate layout design. Microstrip traces are magnified to show their relative size differences.
Figure 11a illustrates the fully assembled optical transceiver package. This includes the CMOS transceiver chip, VCSEL
and photodiode array chips, MT connectorized polymer waveguide arrays and decoupling capacitors, all attached to a
common microprocessor land grid array (LGA) substrate. The polymer waveguide device consists of a 45o mirror on one
end and a modified industry standard multi-terminal (MT) connector on the other. The waveguide array has a core
dimension of 35µmx35µm.
VCSEL/Photodiode array
MT optical connector
Waveguide arrays
Bump
CMOS IC
T-line
Mirror
Microprocessor substrate
Ball grid
Printed circuit
Figure 11. (a) A fully assembled optical transceiver unit (b) a side view of optical coupling of lasers/photodetectors to waveguides
through a 45 degree mirror.
Proc. of SPIE Vol. 6899 68990Z-8
The transmitter includes a VCSEL optical source array and a polymer waveguide array. VCSEL arrays are flip-chip
bonded on the substrate and coupled to polymer waveguides with 45o metal mirrors to direct light at right angle for
transmission through the waveguide. Similarly, the receiver consists of identical optical waveguide assembly to the
transmitter section with a high-speed GaAs photodiode array replacing the VCSEL. Figure 11b shows a side view of the
interfacing between optoelectronics chips and waveguides. The optoelectronic chips are flip-chip bonded with their
apertures face down to allow polymer waveguides to slide under and couple. The total optical loss budget for our system
includes VCSEL/photodiode coupling loss to the 45o mirror on either side of the optical link, propagation loss through
the waveguide, MT connectorization loss and Fresnel losses at the interfaces of connectors. Based on various parameters
of individual optical components, the total optical loss budget for the complete link was calculated to be between 7dB
(best case) and 12dB (worst case)9.
4. MEASUREMENT RESULTS
4.1 Optical and electrical characterization setup
Figure 12 shows a photograph of the electrical/optical measurement setup. Optical transceiver packages are mounted on
high-speed testing sockets on a printed circuit board (PCB). Bias and power were supplied through connectors on the top
surface of the PCB and high speed differential output signals were collected from edge mount connectors. Using a
computer with Labview interface software, a scan chain on chip sent signals to configure parameters in the transmitter
and the receiver. The VCSEL bumps on the LGA substrate were directly probed to acquire electrical signals from the
chip.
Figure 12. A photograph of the experimental set-up for optical transmission tests.
4.2 High-speed electrical and optical characterization
10Gb/s transmitter and receiver
Figure 13 below shows the 10Gb/s optical measurement results of the transmitter and receiver for the two channels
mentioned in section 2.1. These channels are integrated on the main transceiver chip but do not contain circuit blocks
other than the pre-emphasis VCSEL driver and the TIA/LIA pair. For the transmitter measurement in Figure 13a highspeed external electrical PRBS data is received by the on-chip integrated limiting amplifier to drive the pre-emphasis
VCSEL driver. VCSELs were biased with an average current of 7mA.
I
*
1Ml
(a)
(b)
Figure 13. Optical eye diagrams from fully packaged transmitter (a) and receiver (b).
Proc. of SPIE Vol. 6899 68990Z-9
MeasureMcurrent
enimum
main mum
___
I utal mean
peut
Bit Rate(4) 10.00bps? 9.34Gbps? 10.3Gbps? 39
Ee Gmp(4) 143mG
143mG
144 my
24
Figure 14. Received and retimed data from CDR at 10Gb/s driven electrically off-chip.
The resulting optical data was measured by coupling a 1x12 multimode fiber with MT connectors and feeding individual
optical channel outputs to a 12GHz Newfocus optical photoreceiver and 30GHz Agilent DCA oscilloscope. The
measurement has an eye opening of 70ps. The rise and fall times were 30ps and 35ps, respectively with a peak-to-peak
jitter of 30ps. Figure 13b shows the receiver eye opening at 10Gb/s. An eye opening of 60ps and rise and fall times of
25ps and 30ps with a peak-to-peak jitter of 30ps were measured. Figure 14 shows a received and retimed optical data at
10Gb/s from the universal channel containing the CDR in the receive signal path. The eye diagram has a horizontal eye
opening of 60ps and rise and fall times of 35ps and 40ps, respectively. The received eye patterns have some ISI that may
have been caused by losses from the long transmission line for the electrical output signal on the PCB.
20Gb/s transmitter
Figure 15 shows the electrical and optical eye diagrams of the transmitter on the universal channel at 20Gb/s and 18Gb/s
data rate using a 215-1 PRBS data pattern which is generated by the on-chip PRBS generator after the transceiver chip
was packaged. The electrical data was measured by directly probing one of the on-package VCSEL array contact pads of
the pre-emphasis driver with a high-speed GSG coplanar probe and feeding the output to a 30GHz Agilent 86100 series
digital component analyzer (DCA). Before its detection the high-speed signal passes through two package solder bumps
for the CMOS and the VCSEL chips. A microstrip transmission line length of ~4mm connects the two bumps.
Impedance discontinuities were expected at all these points that compromise the signal integrity. Some intersymbol
interference (ISI) and reflection resulting from the signal path on chip and off-chip is apparent in the electrical eye
diagram. However, a good 20Gb/s open electrical eye is observed with a vertical and horizontal eye opening of 175mV
and 35ps, respectively. Also, the measurement showed a rise and fall time of 25ps and 30ps with a peak to peak jitter of
16ps. The optical data was measured by feeding the 20Gb/s electrical signal in Figure 15a directly to the VCSEL using
high-speed coplanar probes and feeding the optical output to a 12GHz Newfocus photoreceiver using a multimode fiber.
Since the signal path now encounters combined parasitics from the package and from the VCSEL, the driver currents
with the pre-emphasis and the associated DC biases were re-optimized to provide the best possible eye at 18Gb/s.
-
)
)
r
tIt
File
I)
Control
O0t1fl3ttOtO
Setup Measure Calibrate
tO.S1fl3ttVtO
(a)
VOtCreVelt
Utilities
ajVOtereVent
Help
:4°t
(b)
Figure 15. Transmitter eye diagrams (a) 20Gb/s Electrical (b) 18Gb/s Optical.
Proc. of SPIE Vol. 6899 68990Z-10
05 NOV 200t 10:35
I
An average drive current of 9mA with 2.8V VCSEL external bias was used to provide an average optical power of 2mW
to the photoreceiver. An 18Gb/s open optical eye was observed with a vertical and horizontal eye opening of 70mV and
30ps (60% of total eye), respectively. The rise and fall time were 25ps and 30ps with an increased peak-to-peak jitter of
23ps. The increase in the jitter and noise around the eye might be associated with the optical device relative intensity
noise. Reflections and bandwidth limitations of the 12GHz photoreceiver also contributed in reducing the eye opening.
SUMMARY
This paper introduced the interconnect bottleneck problems in the PC and server space using existing I/O architectures
and described the role optics could play in overcoming these because of its good attributes for signal propagation. As a
demonstration of the viability of optics for these computing applications a high-speed optical CMOS based transceiver
package has been developed. With this hybrid package 20Gb/s electrical and 18Gb/s optical eye diagrams for LC-PLL
based VCSEL transmitters has been measured. 10Gb/s Fully assembled transmitters and receivers with clock data
recovery were also demonstrated.
ACKNOWLEDGMENTS
The authors would like to thank Douglas Stunkard and Julie Mckenney for chip layout design and Bent Danielson for
test board layout design.
REFERENCES
1.
J. Held, J. Bautista, and S. Koehl, “From a few core to Many: A Tera-scale Computing Research Overview,”
Research at Intel white paper, download.intel.com/research/platform/terascale/terascale_overview_paper.pdf,
2006.
2.
K. Kuhn, et al.,, “A 90nm Communication technology featuring SiGe HBT transistors, RF CMOS, Precision RL-C RF elements and 1um2 6-T SRAM cell”, IEDM 2002, Dig. Tech. Papers, Dec 2002, pp. 73-76.
3.
D. Kucharski, Y. Kwark, D. Kuchta, D. Guckenberger, K. Kornegay, M. Tan, C-K. Lin and A. Tandon, “A
20Gb/s VCSEL driver with pre-emphasis and regulated output impedance in 0.13um CMOS” in Proc. ISSCC,
Feb. 2005, pp. 222-223.
4.
C. Kromer, C. Sialm, C. Berger, T. Morf, M. L Schmatz, F. Ellinger, D. Erni, G. L. Bona and H. Jackel, “A
100mW 4x10Gb/s Transceiver in 80nm CMOS for High-Density Optical Interconnects” IEEE JSSC, vol. 40, no
12, pp. 2667-2679.
5.
H. Muthali, T. Thomas and I. Young, “A CMOS 10Gb/s SONET Transceiver,” IEEE JSSC, vol. 39, No. 7, pp.
1026-1033, July 2004.
6.
J. J Morkuni, P. V. Vena, A. V. Harton, K. W. Wyatt and S. M. Kang, “Spatially Independent VCSEL Models
for the Simulation of Diffusive Turn-Off Transients,” J. Lightwave. Technol., vol. 17, no. 1, pp. 95-102, Jan.
1999.
A. Kern, A. Chandrakasan and I. Young, “18Gb/s Optical IO: VCSEL Driver and TIA in 90nm CMOS,” VLSI
Circuits, 2007 IEEE Symposium, pp. 276-277, 14-16 June 2007.
7.
8.
E. Mohammed, T. P. Thomas, D. Lu, H. Braunisch, S. Towle, B. C. Barnett, I. A. Young and G. Vandentop
“Optical I/O Technology for Digital VLSI” in Proc. SPIE Photonic West Conference, vol. 5358, June 10, 2004.
9.
E. Mohammed, A. Alduino, T. Thomas, H. Braunisch, D. Lu, J. Heck, A. Liu, I. Young, B. Barnett, G.
Vandentop, R. Mooney, “Optical Interconnect System Integration for Ultra-Short-Reach Applications,” Intel
Technology Journal, Vol. 8, no. 2, pp. 115-127, May 2004.
Proc. of SPIE Vol. 6899 68990Z-11