Title: A 20Gbps Serial Link for HighLoss Channels
1A 20-Gbps Serial Link for High-Loss Channels
- Sameh Ibrahim
- Behzad Razavi
2Outline
- Definition of the problem
- Architectural solutions
- Multi-Tone
- DFE
- Circuit solutions
- Layout solutions
- Measurement results
- Conclusions and future work
3Motivation
- Many applications require data transmission in
the Gb/s range. - Manufacturers are seeking data rates in the range
of 20 Gb/s and beyond. - At such high speeds, many design challenges exist.
4Frequency-Dependent Loss
- Conductor and Dielectric Loss
5Cross-Talk
6Reflections
7A Practical Backplane Example
8Measured FR4 Traces Losses
9Equalizer Classifications
- Tx vs. Rx Equalization
- Digital vs. Analog
- Linear vs. Non-Linear
TX Eq
RX Eq
Transmission Channel
ADC
FFE
Slicer
FBE
DFE
Analog Equalizer
Sense Amplifier
Adaptation
10Equalizers of Interest
- Linear but with multi-tone signaling
- Non-linear DFE in analog domain
11Outline
- Definition of the problem
- Architectural solutions
- Multi-Tone
- DFE
- Circuit solutions
- Layout solutions
- Measurement results
- Current and future work
(ISSCS July 2009)
12Multi-Tone Signaling
fc1
Transmitter
S/P
LPF
Mod
fc2
LPF
Mod
Binary Data
fcN
LPF
Mod
Channel
fc1
P/S
Demod
LPF
fc2
Demod
LPF
Binary Data
fcN
Demod
LPF
Receiver
13Advantages of Multi-Tone Signaling
- In every subchannel, processing can be done at a
lower rate. - Equalization can be simple or abandoned.
- Longer symbol time improves immunity to impulse
noise. - Allows discrete pre-emphasis. (discussed later)
- No need for high-speed mux/demux, PLLs or CDR
circuits.
14Challenges
- Complexity
- Modulation type implementation.
- RF operations quadrature mixers oscillators.
- PLLs and CDR circuits.
- DACs and ADCs.
- Circuit Design
- TX RX linearity.
- Problem of harmonics.
- I/Q mismatches.
- Effect of adjacent channels.
- Tilt correction.
15Shall we Proceed with Multi-tone?
- While the equalization task is simple in
multi-tone signaling, new design challenges are
introduced. - The complexity of the system prevents its use
whenever a binary solution is possible. - For very high data rates or multi-band
transmission channels, multi-tone signaling is an
attractive choice.
16Outline
- Definition of the problem
- Architectural solutions
- Multi-Tone
- DFE
- Circuit solutions
- Layout solutions
- Measurement results
- Current and future work
(Submitted ISSCC 2010)
17DFE Main Equation
DFE
xrec
yeq
Slicer
xtr(n-1),xtr(n)
_
xfb
Loop Delay lt UI 50ps
xtr(n-1)
FF
A
xrec(n)hoxtr(n)h1xtr(n-1)
xfb(n)Axtr(n-1)
DFE
yeq(n)hoxtr(n)h1xtr(n-1)- Axtr(n-1)
FFE
If A h1 yeq(n)hoxtr(n)
18Advantages and Disadvantages of DFEs
Equalized Data
ADC
FFE
Slicer
Received Data
FBE
DFE
- Best performance
- Corrects both loss and reflections
- Dont increase cross-talk
- High power (especially digital)
- Timing constraints
19DFE Implementations
- Relaxes time constraints.
- Reduces the loading at the summing node.
- Increases complexity.
20Prior Art
- Few CMOS solutions for rates near 20 Gb/s have
been reported.
21Direct Full-Rate DFE
R
R
R
Received Data
Equalized Data
2/R
h1
h6
h2
tc2q tsetup tFB lt UI 50ps
22Unrolled Full-Rate DFE
R
R
R
2/R
h1
h6
h2
Received Data
Equalized Data
R
R
R
2/R
h1
h6
h2
tc2q tsetup tc2q,MUX lt UI 50ps
23Direct Half-Rate DFE
R
R
R
tc2q tsetup tFB lt UI 50ps
Even Data
2/R
h1
h6
h2
Received Data
R
R
R
Odd Data
2/R
h1
h6
h2
24Multiplexed Half-Rate DFE
R
R
Even Data
Received Data
2/R
h1
h6
h2
tc2q tsetup tprop,MUX tFB lt UI 50ps
Can be merged
Odd Data
Payne et al., JSSC Dec 2005
25Unrolled Half-Rate DFE
R
R
L
L
2/R
h1
h6
h2
Even Data
R
R
R
L
L
2/R
h1
Received Data
tc2q tsetup tc2q,MUX lt UI 50ps
h6
h2
R
R
L
L
2/R
h1
h6
h2
Odd Data
R
R
R
L
L
2/R
Bulzacchelli et al., JSSC Dec 2006
h1
h6
h2
26Multiplexed Unrolled Half-Rate DFE (1)
R
R
Even Data
Received Data
2/R
h1
h6
h2
Odd Data
27Multiplexed Unrolled Half-Rate DFE (2)
R
R
Received Data
Even Data
2/R
h6
h2
h1
Odd Data
28Multiplexed Unrolled Half-Rate DFE (3)
R
R
Even Data
2/R
h6
h2
h1
Received Data
R
R
L
L
2/R
Odd Data
h1
h6
h2
29Multiplexed Unrolled Half-Rate DFE (4)
R
R
Even Data
2/R
h1
h6
h2
Received Data
R
R
L
L
2/R
h1
Odd Data
h6
h2
tc2q tsetup tc2q,MUX tprop,MUX lt UI 50ps
30Merged MUXs Block
Q1
Ze
Q
Q1
Q0
Q
Q0
Zo
Ze
Clk
Zo
Clk
31Multiplexed Unrolled Half-Rate DFE (5)
R
R
Even Data
2/R
h1
h6
h2
Received Data
R
R
L
L
2/R
h1
Odd Data
h6
h2
tc2q tsetup td2q,MM lt UI 50ps
32Multiplexed Unrolled Half-Rate DFE (6)
Even Data
R
R
2/R
h1
h6
h2
Received Data
R
R
L
L
2/R
h1
h6
h2
Odd Data
td2q td2q,MM lt UI 50ps
33Multiplexed Unrolled Half-Rate DFE (7)
Even Data
R
R
2/R
h1
Received Data
h6
h2
2/R
L
L
h1
R
R
Odd Data
34Multiplexed Unrolled Half-Rate DFE (8)
Even Data
R
R
A
1/R
h1
Received Data
h6
h2
A
1/R
L
L
h1
R
R
Odd Data
Clk 1
Clk 2
35Final Proposed Equalizer
Even Data
R
R
A
1/R
h1
Received Data
FFE
h6
h2
A
1/R
L
L
h1
R
R
Odd Data
Clk 1
Clk 2
36Final Proposed 1-Tap Equalizer
37Proposed Architecture Advantages (1)
- Half-rate clocking with loop unrolling to achieve
better timing constraints. - Amplifiers added in speculation paths to increase
the swing applied to the MUX and allow faster
current steering . - Multiplexing the two branches reduces the number
of summing nodes saving power and are. - A latch rather than a flipflop can be used as the
feedback memory element improving the speed of
the critical path.
38Proposed Architecture Advantages (2)
- Using stacked MUXs saves the delay of one MUX in
the critical path. - Inductive peaking is used to extend the bandwidth
of all the blocks in the critical path. - Half-rate clocking enables the use of
higher-swing clocks and better phase accuracy. - Half-rate clocking also enables the use of
lower-power CDRs.
39Outline
- Definition of the problem
- Architectural solutions
- Multi-Tone
- DFE
- Circuit solutions
- Layout solutions
- Measurement results
- Current and future work
40Linear Equalizer
41Linear Equalizer Simulation Results
42DFE Circuit Realization
- CML latches and MUX with class AB biasing and
inductive peaking. - R1 is added to adjust the DC for direct coupling
to the stacked MUX block.
43Outline
- Definition of the problem
- Architectural solutions
- Multi-Tone
- DFE
- Circuit solutions
- Layout solutions
- Measurement results
- Current and future work
44Layout
Input
Output
Clock
Core
Diff. TLs
800 µm
Serial Interface Bus
DC Pads and ESDs
1.4 mm
45SIB
- 13 bits
- h1Coefficient 61(sign) bits
- Linear Equalizer Control 3 bits
- Clock Buffer Control 3 bits
46Core
MR
Clock Buffer
Delay Stage
B
Delay Stage
CC
Output Buffer
CC
Summing Node Inductors
MR Matching Resistors CC Coupling
Capacitors C MUX Current Multiplexers B
Biasing
Feedback Latch
Amplifier
350 µm
Merged MUXs
Latches
C MUX
Gm
Amplifier
Feedback Latch
MR
Summing Node Inductors
Output Buffer
300 µm
B
Current Sources
47Inductors
- Multi-layer stacked inductors to reduce area
- M3, M6 and M9 used to reduce capacitance
- Modeled by HFSS and Matlab
48Die Photos
49Outline
- Definition of the problem
- Architectural solutions
- Multi-Tone
- DFE
- Circuit solutions
- Layout solutions
- Measurement results
- Current and future work
50PCB Board
- 2 layers
- No components or exposed traces in bottom layer
- Enough space for probing
3.50 inch
3 inch
51PCB Photo
52Wire Bonding
53Test Setup
Centellax MS4S1M
Centellax TG2P1-A
FR4 Trace
ZRON8
Centellax UXC20P
/2
Centellax TG1B1-A
10 GHz
To Scope Trigger
Centellax MD1S4M
ZVA-183
Agilent E8257D
Agilent E8257D
10 GHz
54Setup Photo
55PRBS Generator Output
horizontal scale 20 ps/div vertical scale 100
mV/div
- 20-Gb/s data
- 10-ps p-p Jitter
- Clock adds another 7-ps p-p jitter
56Channel Frequency Response
- 18-inch FR4 plus 2 cables
- 24 dB of loss at 10 GHz
57DMUX Output
horizontal scale 20 ps/div vertical scale 50
mV/div
- 20-Gb/s input give 10-Gb/s output
- After 18-inch channel with 24 dB of loss
- Limited bandwith of the output buffer adds to the
eye closure
58Bathtub Curves
59Performance Summary
60Conclusions
- Conclusions
- A 1-Tap DFE running _at_ 20-Gb/s
- A DFE capable of equalizing 15 dB of loss with
appropriate BER and eye opening - Lowest figure-of-merit reported in literature
(less than half of any FOM reported)