Modeling of SerDes channels using IBIS-AMI ... Standard API makes co-simulation between two ... Equalization is optimize
Improving IBIS-AMI Model Accuracy: Model-to-Model and Model-to-Lab Correlation Case Studies Dong Yang1, Yunong Gan1, Vivek Telang1, Magesh Valliappan1, Fred S. Tang1, Todd Westerhoff2, and Fanyi Rao3 1Broadcom
Corporation
5300 California Ave, Irvine, CA 92617 3Agilent
2Signal
Integrity Software, Inc. (Sisoft)
6 Clock Tower Place, #250, Maynard, MA 01754
Technologies, Inc.
5301 Stevens Creek Blvd, Santa Clara CA 95051
Agenda • Introduction: – Modeling of SerDes channels using IBIS-AMI – LinkEye® – Broadcom’s in-house simulation tool • Case study 1: model-to-model correlation: – 11.5G SerDes IBIS-AMI model • Case study 2: model-to-lab correlation: – 10G 40nm XFI transmitter correlation – 10G 40nm XFI transmitter + receiver correlation – SFI-5.1 to OTU3 (2x23Gbps D-QPSK) MUX transmitter correlation • Discussion and conclusions
SerDes Channel
Why AMI? – Sophisticated equalization (filter, FFE, DFE) and clock/data recovery techniques – SerDes models are traditionally available as transistor-level models: • Low BER (10-12 – 10-18) requires impractically long simulations, large simulation capacity required.
• Difficult to be ported over different EDA platforms. – Ability of simulating Tx and Rx from different IP vendors – IP protection
IBIS – AMI Model Tx AMI Tx DSP
Rx AMI Tx Analog Front End
Channel Model
Rx Analog Front End
Rx DSP
AMI: delivered as dynamically linked library (DLL): – Executable “black box” protects vendor IP. – Flexible and accurate model controlled by IP vendor. – Standard API makes co-simulation between two different vendors possible.
Agenda • Introduction: – LinkEye® – Broadcom’s in-house simulation tool
LinkEye: Simulation Model Tx Driver
TP1
TP2
TP3
TP4
TP5
B1
TP6 Rx Slicer
B2 HMD (wc)
HTx
Transmitter
S11 S12 S S * 21 22
Channel
S11 S12 S S * 21 22
• S-domain measurements to ABCD domain • Multiply together • ABCD domain to S-Domain • Inverse FFT to obtain impulse response
S11 S12 S S 21 22
HRx
Receiver
IFFT
Normalized Amplitude
A
TP0
Time (ns)
LinkEye: Simulation Methodology • MatLab simulations: – Pulse response “frequency-domain” analysis – MMSE optimization
• Performance evaluation based on detailed, worst-case error probabilities (not simple Gaussian assumption) • On-chip impairments included: – – – – –
Clock jitter Offsets Front-end noise Detailed analog-circuit models Detailed equalizer-implementation penalties
• Target BER = 10-18 (corresponds to target SNR = 19 dB)
LinkEye: Worst-Casing • Data patterns: – ISI added destructively with worst-case bit sequences. – Crosstalk added destructively with worst-case bit sequences.
• Phase alignment: – Reflections added destructively with worst-case phase. – Crosstalk phase assumed to be worst case aligned with signal.
• Launch amplitude: – Link partner is skewed to low end of allowable range. – Crosstalk aggressor is skewed to high end of allowable range.
• Launch rise time: – Link partner is skewed to slow end of allowable range. – Crosstalk aggressor is skewed to fast end of allowable range.
• All analog filters use worst-case PVT parameters. • Actual package and driver models are used.
LinkEye: Equalizer Optimization • Transmit preemphasis equalization is set at start-up: – Start-up protocol can be used to automatically set Tx preemphasis settings.
• Receiver equalizer adapts continually: – Compensates for temperature, environment changes.
• Optimization code: – For each Tx preemphasis setting, compute optimal receive equalizer. – Picks best combination globally.
• Equalization is optimized using Minimum-Mean-SquaredError (MMSE) techniques combined with low-probability BER estimation.
Agenda • Introduction:
• Case study 1: model-to-model correlation: – 11.5G SerDes IBIS-AMI model
11.5G SerDes AMI Model Backplane Channels Tx AMI Model Tx Analog Front-End (On-Die S-Parameter)
-a
1
Rx AMI Model Peaking Filter VGA Rx Analog Front-End (On-Die S-Parameter)
-b
Rx Slicer
S DFE
Tx Preemphasis
• On-die S-parameter is introduced to characterize the analog front ends of Tx and Rx AMI models. • Rx DSP includes peaking filter (PF), VGA, and a DFE. • Some backplane channels are used in AMI-to-LinkEye correlation.
AMI-to-LinkEye Correlation Simulation Comparison Before Optimization Channel ID
AMI (Statistical Sim.)
LinkEye
P.K.F.
VGA
DFE1
BER
P.K.F.
VGA
DFE1
BER
1
7
15
0.2601
3.00E-39
4
16
0.358
1.00E-69
2
7
15
0.2776
1.00E-39
5
16
0.354
1.00E-65
3
12
17
0.4727
2.00E-36
9
19
0.541
1.00E-45
4
15
22
0.7114
2.00E-28
15
22
0.741
1.00E-30
5
8
16
0.439
5.06E-09
5
17
0.475
1.00E-10
6
10
17
0.487
8.25E-10
6
18
0.527
1.00E-10
7
15
21
0.665
2.85E-12
13
20
0.650
1.00E-13
8
9
15
0.366
1.99E-19
6
17
0.439
1.00E-18
9
7
15
0.302
2.57E-19
5
16
0.384
1.00E-22
10
15
22
0.697
1.25E-11
14
21
0.687
1.00E-12
11
15
19
0.553
4.62E-21
12
20
0.592
1.00E-22
12
9
15
0.366
1.99E-19
6
17
0.439
1.00E-18
Poor correlations between first version of AMI model to LinkEye® model.
Comparison of AMI and LinkEye Models Ideal Impulse
A1
Tx AMI (FIR)
A2
Tx Driver
Reshape
S-para:
NA
Tx Pkg
A3
Channel
A4
Rx Pkg
Reshape
Reshape
Rx Load
Reshape
S_AMI_txp S_AMI_channel S_AMI_rxp
S_AMI_txd
A5
A6 Rx AMI (PKF, VGA, DFE)
EDA Tool’s CDR & Decision Circuit
Reshape
AMI
S_AMI_rxl
• S_AMI_txd = S_LE_txd; S_AMI_txp = S_LE_txp • S_AMI_rxl = S_LE_rxl; S_AMI_rxp = S_LE_rxp
L1 Tx Driver
L2 Tx Pkg
Reshape
S-para:
S_LE_txd
Reshape S_LE_txp
L3 Channel
Reshape S_LE_channel
L4 Rx Pkg
Reshape S_LE_rxp
L5 Rx Load
L7
L6 Impedance calibration
Rx EQ Optimization (MatLab Code)
LinkEye CDR & Decision Circuit
Reshape S_LE_rxl
LinkEye
Model-to-Model Correlation Strategy Step 1. Pulse response at Rx input for AMI and LinkEye AMI
Tx Driver (Pkg included)
Channel
Rx Load (Pkg included)
S_AMI_tx
S_AMI_channel
S_AMI_rx
Ideal Pulse
Step 2. Rx block calculation for AMI and LinkEye Rx AMI (PKF, VGA, DFE)
Force the same input
Compare Rx EQ Optimization (MatLab Code)
LinkEye
Rx Input
matches LinkEye
Pulse Responses at Rx Input
Sanity Check #1: Impulse-to-Pulse Conversion Test 1: IR -> PR
1) Starting with EDA tool’s IR, calculate PR using LinkEye code (IREDA*ideal_pulse). 2) Add some offset, then plot with EDA tool’s PR - they are exactly matched!
Sanity Check #2: Pulse-to-Impulse Conversion Test 2: PR -> IR
1) Starting from EDA tool’s PR, calculate IR using LinkEye code. 2) Plot it with EDA tool’s IR - they are exactly matched!
Optimization of On-Die S-Parameters IR (tx driver + tx pkg + channel + rx pkg + rx load) AMI (Red) vs. LinkEye (Blue)
Before the optimization
After the optimization
Rx Block Calculation Comparison EQ parameters match! AMI (PF, VGA, first two DFE taps): 13 20 0.6494 -0.0194 LinkEye (PF, VGA, first two DFE taps): 13 20 0.6497 -0.0195
A5
A6
Rx AMI (PKF, VGA, DFE)
Output pulses match!
Force AMI input = LinkEye input
LinkEye @ L7 AMI @ A6
LinkEye L6
Rx EQ Optimization (MatLab Code)
L7
EQ parameters match
+ Output pulses match
AMI Rx LinkEye
Simulation Results After Optimization Sample Eye Plot: LinkEye vs. AMI LinkEye
AMI Simulation
Simulation Results: LinkEye vs. AMI SimID Precursor Main Post-cursor
LinkEye Simulation Results AMI Simulation Results PF_Low PF VGA DFE Tap1 DFE Tap2 Eye Width Eye Height PF_Low PF VGA DFE Tap1 DFE Tap2 Eye Width Eye Height
316 316
3 3
0 -0.125
0.8 0.8
0 0
10 11 8 13
0.852 0.842
0.188 0.193
0.23UI 0.25UI
11% 12.40%
3 3
9 8
11 13
0.859 0.847
0.192 0.185
0.25UI 0.27UI
12.40% 14.00%
More Simulation Cases 11.5G LinkEye vs. EDA Tool Correlation (Typ. PVT) Channel IL @5.75GHz ID (dB) 1
7.8
2
12.9
3
14.3
4
15.0
5
16.8
6
17.6
Veye (mv)
Heye (ps)
Tx FIR
VGA PF_low
PF
DFE1
122
38.4
27
107
43.9
[-0.125 0.675 0.000]
1
3
0.541
27
1
3
0.528
113
38.4
[-0.125 0.675 0.000]
27
4
5
0.644
93
43.0
28
3
5
0.649
88
30.3
[-0.125 0.675 0.000]
28
5
5
0.669
90
39.5
29
5
5
0.692
114
38.4
[-0.125 0.675 0.000]
31
3
5
0.710
119
46.5
31
3
5
0.706
101
34.4
[-0.125 0.675 0.000]
29
6
6
0.717
79
43.4
29
6
6
0.740
115
38.4
[-0.125 0.675 0.000]
31
5
6
0.750
112
47.3
32
4
6
0.773
Agenda
• Case study 2: model-to-lab correlation: – 10G 40nm XFI transmitter correlation
10G 40nm XFI IBIS-AMI Model (Tx Correlation)
Measurement Setup
Tx Correlation BER = 1e-6 Eye Height (inner) @1e-6 (mv)
Eye Height (outer) @1e-6 (mv)
Eye Width (inner) @1e-6 (ps)
Measured
Simulation
Measured
Simulation
Measured
Simulation
8
177
169
359
381
77
73
14
85
79
229
229
69
70
Trace length (inch)
Main tap
Postcursor
Case 1
12
21
Case 2
31
17
Test ID
BER = 1e-12 Eye Height (inner) @1e-12 (mv)
Eye Height (outer) @1e-12 (mv)
Eye Width (inner) @1e-12 (ps)
Measured
Simulation
Measured
Simulation
Measured
Simulation
8
148
146
388
399
65
64
14
69
69
250
240
60
63
Trace length (inch)
Main tap
Postcursor
Case 1
12
21
Case 2
31
17
Test ID
Test Case 1: Trace = 12”, Main_tap = 21, and Post-Cursor = 8
Test Case 2: Trace = 31”, Main_tap = 17, and Post-Cursor = 14
Agenda
• Case study 2: model-to-lab correlation: – 10G 40nm XFI transmitter + receiver correlation
10G 40nm XFI IBIS-AMI Model (Tx+Rx Correlation)
Measurement Setup
Tx+Rx Correlation Test ID
Trace (inch)
Tx Main Tap
Tx Post Tap
Rx Peaking Filter
Measured BER
Simulated BER
Case 1
31+12
24
0
14
3.0E-7
7.5E-5
Case 2
31+24
24
15
8
3.4E-12
1.8E-11
Case 3
31+12
24
0
8