Title: CMOS Optical Links ORS
1CMOS Optical LinksORS
- Samuel Palermo, Azita Emami-Neyestanak, Hae-Chang
Lee, and Mark Horowitz - Stanford University
2 Overview
- High Speed Links
- Introduction to electrical links
- Limitation of electrical links
- Benefits of optical links
- Optical link architecture
- Optical Transmitters
- VCSEL and Modulator drivers
- Trends in data rate, power, and supply voltage
- Optical Receivers
- Comparison of different front ends
- Projection of power and area of RX in the future
- New receiver for optical routers
3High Speed Electrical Links
- Necessary to equalize the growing disparity
between on-chip computation and chip-to-chip
communication bandwidth - Components
- High-bandwidth transceiver (TX, RX)
- Terminated channel
- Precise clock generation and recovery
4Limitations of Electrical Links (1)
- Maximum on-chip clock frequency that can be
propagated without swing attenuation - Clock period limit ? 6 8 FO4 inverter delays
- 0.25? CMOS ? 750 1000ps ? 1 1.3GHz
5Improving Data Rate
- Multilevel signaling
- Requires small proportional noise sources
(reflections cross-talk) due to reduction in
SNR - Time-interleaved multiplexing
- Requires multi-phase clock generation with low
static phase offset and jitter - Improvement is still less than 10X
6Limitations of Electrical Links (2)
- Limited bandwidthdistance product of wires
- Proportional noise sources
- Reflections
- Cross-talk
- Power Consumption 30mW/Gb/s
(D. Miller and H. Ozaktas, J. Parallel Distrib.
Comput., 1997)
7Benefits of Optical Links
- Better Channel
- Longer Distances (SMF)
- Less attenuation per distance (0.2dB/km)
- Almost zero frequency dependant loss
- Dispersion Limited
- Chromatic (5ps/nm/km)
- Lower Power
- Less attenuation
- Less Noise
- No crosstalk between fibers
- No reflections
8Multi-Gbps CMOS Optical Link Architecture
9VCSEL Optical Source
- Optical power a linear function of laser diode
current once biased above threshold - High contrast ratio can be achieved by biasing
close to threshold - High frequency operation requires biasing the
VCSEL above threshold to avoid turn-on delay - Optical frequency response proportional to
10VCSEL Electrical Model
Small Signal Model
- Device electrical bandwidth dominated by bias
dependent RJCJ - RJ 30 - 100?
- CJ 0.5 1pF
- ? 40 70ps
- Total series resistance between 50 - 100? means
low voltage swing for power levels appropriate
for short optical interconnects - Parasitic inductance causes current ringing
reduced voltage headroom - Device requires current driver with adequate
range to handle variations in threshold current
and slope efficiency
11TX VCSEL Driver Output
- Differential current steering driver minimizes
power supply noise - 8 bit control for bias modulation currents
- Current range from 0 5mA with 20?A resolution
- 3.3V output stage to keep current sources
saturated due to VCSEL VF
12TX - VCSEL Driver Simulation Results
- Simulations with IBIASIMOD1mA
- Adequate bandwidth for 5Gbps data
- Area 300? x 520?
13External Modulation with MQWM
- Intrinsic quantum well region in the p-i-n region
is absorbing - Static reverse bias gives a photodetector
- Changing bias shifts the absorption edge - due to
thequantum-confined Stark effect - and
modulation occurs
(figure and data courtesy of G. Keeler, EE,
Stanford)
14MQWM Electrical Model
- Model simplifies to only a lump capacitor
composed of the diode capacitance and parasitic
bonding capacitance - Assuming between 50 200fF total capacitance
- Device requires voltage driver with high output
swing (2 3V) for a contrast ratio of about 3dB
15TX MQWM Driver Output
- Low-swing low-fanout psuedo-nmos mux used for
speed - Low swing amplified to provide 2.5V swing to the
modulator - Propagating fast pulse causes low fanout buffer
chain - Trade-off between number of buffers (jitter
performance) and power consumption due to static
power of psuedo-nmos mux - Modulator bias voltage limited by decoupling
transistors
16 Optical to Electrical Conv.
- Photo-diodes are used for conversion
- Responsivity A/W _at_ l
- Parasitics
- Current noise (shot noise, back ground
illumination)
Vrev
Iop
Cp
In
Electrical model
17Receivers for Optical Links
- Design goals
- High gain, low input noise current, large
bandwidth - Short-haul parallel links
- Dense arrays of receivers
- Low power, area
- Cheap standard process
18Front End Design
- Small R
- High noise
- Large R
- Low BW
t RCp
n2 4KT / R
19Transimpedance Amp
- Low Input impedance
- CG TIA
- Shunt-shunt feed back (R or C)
- t Rf Cp / (A-1)
- Afford larger Rf
- Higher Gain (R)
- Lower thermal noise
Rf
A
A2
Cp
Example shunt-shunt res. feedback
20TIA Design
- Challenges
- High GxBW
- Stability
- Noise/speed/gain trade off
- BW0.7 bit rate (Trade off between ISI and noise)
- Amp. Noise
- Voltage headroom
- Post amplifier/limiter
- Power consumption
21RF CMOS Performance Trend
RF-CMOS Performance Trends P.H. Woerlee et al.
IEEE Trans. Ele. Dev., VOL. 48, NO. 8, Aug 2001
scales
Max stable gain _at_ 2GHz
22RF Performance (Cont.)
Linearity
scales
23RF FOM Trend
24Receivers for VLSI Photonics
Vdet
V1
- Diode clamp front end
- Simple, low power, area
- Needs differential input
- Low sensitivity
V2
-Vdet
25New Receiver
Vdd
di
di0 aIop I0 di1 aIop I1
Iop
Cp
Vin
26Double Sampling, Data Resolution
Vdd
Vin(t1) gt Vin(t0) ad1 1
di
Vin(t)
1 1 0 1
t0 t1 t2
time
d0 d1 d2
27Receivers Block Diagram
1.6 Gbps, 3mW, S11mA
Vdd
f1
Sense Amp
-
Vin
C1
Photodiode
Clk_1
f2
Offset
C2
Sense Amp
-
IDC
Low-Pass Filter
Clk_2
Double-Sampler
Offset
28New Generation with Timing Recovery
- Factor of 5 DeMux at the input
- 5 equally spaced phases
- Frequency and phase adjustment by 2 different
techniques in a dual loop PLL - First technique PClk is shifted by half bit time
from DClk, giving one extra sample for phase
recovery - New technique data samples are used for phase
recovery
5 Gbps, 70 mW
29Optical Routing
- Packet level switching performed in the optics
- Implications
- A) Consecutive packets arrive from different
transmitters - B) Time between packets from the same transmitter
to a given receiver can be extremely long (i.e.
low transition density)
30The Problem
- Different Transmitters have different
phase/frequency w.r.t Receiver - CDR in the receiver must re-acquire phase lock
for each packet - When packets are only a few Kbits, the overhead
can easily run over 30 for conventional CDR - Can solve fast phase acquisition by
- Resynchronize as quickly as possible for each
packet without prior knowledge of transmitters - Measure frequency offset between TX and RX and
guess what the phase will be at the next packet
31CDR for ORS
- Objective
- Calibrate and correct for phase drift due to a
max frequency offset of 100ppm - Reduce this offset to less than 0.1ppm
- 0.01 bit offset in 100k cycles.
- The required technology
- how to acquire/track frequency offset information
in digital format accurately - Digital b/c analog circuits leaks over time
- Identifying and controlling quantization error
32Semi-Digital Dual Loop DLL
332nd Order Digital DLL
34ORS System