Transmission Proposal for 10GBASE-T G. Zimmerman, SolarFlare

Transcription

1 Transmission Proposal for 10GBASE-T G. Zimmerman, SolarFlare G. Zimmerman SolarFlare Communications 1

2 Supporters Rick Rabinovich, Spirent Communications Dan Dove, HP Joel Goergen, Force 10 Networks Chris DiMinico, MC Communications Mike Bennett, Lawrence Berkeley Labs Michael Laudon, Force 10 Networks G. Zimmerman SolarFlare Communications 2

3 Outline Overview Baud Rate: Info. Bits/Baud Equalization FEC/Trellis Coding Launch Power Backoff G. Zimmerman SolarFlare Communications 3

4 Overview: Key Choices to Make Line coding Starts with baud rate (bandwidth) Exact # levels of PAM tied to FEC choice May requires overhead for MAC control symbols Depends on coding FEC choice & partition Includes both line coding & partition Launch voltage Power consumption/noise immunity tradeoff EMI constraint Power backoff for short lines G. Zimmerman SolarFlare Communications 4

5 Overview: Elements Considered Channel models 55m Class E Objective: Cabling ad hoc IL, NEXT, FEXT & RL models ( T_Cat6_Model.zip) Class E ad hoc ANEXT model, Class E ISO proposal (15 db/decade) 100m Class E+ Objective: Class E ad hoc IL, NEXT, FEXT & RL models Proposals from TR42, ISO, and 3 rd parties EMI models EMI radiative transfer function derived from measurements presented to IEEE 10GBASE-T Study Group Component effects Magnetics bandwidths Timing recovery effects Info bits/baud determines baud rate Based on Optimal DFE signal processing G. Zimmerman SolarFlare Communications 5

6 Baud Rate: Info bits/baud (/pair) Determines necessary & used bandwidth Performance, Power & EMI Constrained DFE systems generally have a unique optimum Performance vs. baud rate on DFE channels is not identical to AWGN channels Rate loss is channel dependent (Rate loss in DFEs under pinch off conditions: ref. T1E1.4/97-241) Optimal DFE Margin (Salz) normalized to bits/baud: Uncoded Margin = -10*log10(Salz_MSE)- Capacity_SNR+12.27dB Capacity SNR = 10*log10(2^(2*bits/baud/pair) 1) db G. Zimmerman SolarFlare Communications 6

7 Baud Rate: 55m Class E Ad Hoc Model Very shallow optimum ANEXT Model exhibits <10dB/decade ANEXT slope -60 Received Residual PSDs 0.5 Uncoded DFE Margin vs. bits/baud/pair Uncoded DFEMARGIN dbm/hz signal out rem next rem alien next -180 rem fext rem echo rem awgn Uncoded Margin (db) Info bits/baud/pair G. Zimmerman SolarFlare Communications 7

8 Baud Rate: 55m Class E with 15dB / decade ANEXT model Optimum shifts towards 3 bits/baud & steepens ANEXT Model based on presentations Conforms with data(hayes_1_0303.pdf, abughalazeh_1_0903.pdf) ANEXT Loss = 47-15log10(f/100) db (limit line adj) Received Residual PSDs Uncoded DFE Margin vs. bits/baud/pair Uncoded DFEMARGIN dbm/hz signal out rem next rem alien next rem fext rem echo rem awgn Uncoded Margin (db) Info bits/baud/pair G. Zimmerman SolarFlare Communications 8

9 Baud Rate: 100m Class E+ Example DFE Margin vs. info bits/baud strongly favors lower baud rates ANEXT Loss = 60-15*log10(f/100) + 2.5dB (limit line adj.) -80 Received Residual PSDs -1 Uncoded DFE Margin vs. bits/baud/pair Uncoded DFEMARGIN dbm/hz signal out rem next -180 rem alien next rem fext rem echo rem awgn Uncoded Margin (db) Info bits/baud/pair G. Zimmerman SolarFlare Communications 9

10 Baud Rate: EMI Used Field Radiated EMI measurements to estimate transmit PSD within FCC Class A Issues emerge above 500 MHz Transmit PSD with 3 db Margin to FCC Class A EMI (from measured emissions) Transmit PSD (dbm/hz) Frequency (MHz) ref: G. cohen_1_0903.pdf Zimmerman SolarFlare Communications 10

11 Other Components Magnetics performance falls off beyond 500 MHz Adversely effects noise susceptibility & EMI in addition to received SNR ref: dihn_1_0104.pdf Timing jitter degrades SNR as baud rate increases (10*log10(fb1/fb2) relative loss in ADC quant noise PSD) G. Zimmerman SolarFlare Communications 11

12 Baud Rate: Conclusions Choice of info bits/baud (baud rate) is a function of tradeoffs in: Long-line Performance ANEXT Robustness Meeting EMI Other component effects (e.g., Magnetics, timing) 3 bits/baud/pair is within 1 db of optimum point for DFE SNR for all cases, and closer on hard cases 3 bits/baud/pair allows transmit PSD to roll off before 500 MHz Meets EMI, aligns with magnetics rolloff G. Zimmerman SolarFlare Communications 12

13 Equalization: Tomlinson-Harashima Alleviates DFE error propagation in coded systems Cost is large amplitude dither element added to signal Transmit power penalty is small for large # PAM levels Problems: Dither couples through NEXT, FEXT & Echo paths High PAR & extra dynamic range increases complexity Incompatible with shaping gain Requires tight circuit timing loop for feedback filter +/- L (PAM) +/- 2L Dither H txp H chan H ffe mod L Transmit Precoder Filter Wireline Channel Feed-Forward Equalizer H dfe TH Precoder TH Precoded Channel Equalizer Block Diagram G. Zimmerman SolarFlare Communications 13

14 Equalization: Precoded DFE Adaptive Linear precoding can shape DFE response to minimize error propagation Small transmit power penalty for preemphasis 10GBASE-T not generally transmit power limited Can be combined with other transmit filtering Can be combined with constellation shaping gain Feedforward structures minimize circuit timing issues H txp H chan H ffe Transmit Precoder Filter Wireline Channel Feed-Forward Equalizer H dfe Decision Feedback Equalizer DFE-based Channel Equalizer Block Diagram G. Zimmerman SolarFlare Communications 14

15 Equalization: Precoded Precoder coefficients trained at startup to adapt to varying line lengths Max DFE feedback coefficient can be constrained <.25 DFE can be shaped to avoid catastrophic error propagation Precoded DFE Simulated Symbol Error Rates for Uncoded PAM Symbol Error Rate Linear Tx Precoding DFE with Correct Decisions DFE Theory Signal-to-Noise Ratio (db) Transmit PSD and EMI with Precoded Filter 100 meter channel 50 meter span Cat 5e F CC Cla s s A limit FCC Class B limit x 10 8 G. Zimmerman SolarFlare Communications 15

16 FEC/Trellis coding: Latency Applications show need for lower latency codes Distributed computing, clustering require capability for low latency operation Includes propagation, code and signal processing latency Long lines mask PHY latency (propagation delay) Generic Ethernet places no hard requirement on 10GBT Legacy of the fact that 802.3ae was engineered for multi-km links (light time > 5 usec) Previous Ethernet has not stated latency as a requirement High latency codes PERMANENTLY bar technical innovation from achieving low latency operation Additional coding gain can be achieved by layering an outer code, if necessary, on long lines without impairing minimum PHY latency on shorter lines G. Zimmerman SolarFlare Communications 16

17 Code Proposal: 4D-4W 4W-Trellis Code 4D (across pairs) PAM-10 with 4-way time-interleave and constellation shaping Advantages Meets 3 bits/baud information rate Encodes control symbols into modulation, avoiding rate loss Provides for minimal latency operation (<<.25usec) Provides for constellation shaping gain (0.64 db) 4-way interleave allows lower-rate decoder clocking Interleave mitigates noise correlation effects Interleave mitigates error propagation effects Low complexity hardware encoding & decoding Allows concatenation for layering block FEC if desired for improved impulsive noise or long line performance G. Zimmerman SolarFlare Communications 17

18 Line Code Proposal: 4D-4W 4W-PAM10 8st 4D Ungerboeck code used in 1000BASE-T 2^13 possible encoded symbols 10,000 constellation points Remaining 1808 points can be used for control symbols 4 Way time interleaving, code is 4D across pairs Balanced constellation No polarity scrambler required Shaped constellation (0.64 db shaping gain) Table 1: 1D PAM Level Rate of Occurrence in the 4D Mapping (8192 points) G. Zimmerman SolarFlare Communications 18

19 4D-4Way 4Way PAM-10 Code Performance 10 0 Coded PAM10 simulation SER 10-2 Coded PAM10 Theory SER Coded PAM10 Theory BER Slicer Input SNR (db) 26.2 db G. Zimmerman SolarFlare Communications 19

20 Launch Power Tradeoffs Launch power < 10 dbm due to EMI constraints Long line launch power > 6 dbm due to 1000BASE-T ANEXT constraints Negotiated launch power backoff Widely used in deployed DSL standards to mitigate asymmetric link near/far problem Lines less than 50m Negotiated at Startup, based on SNR and/or attenuation Minimum backoffs to be specified in the standard G. Zimmerman SolarFlare Communications 20

21 Baud Rate Proposal Motion #1: That 10GBASE-T baseline baud rates consistent with 3 information bits/baud/pair G. Zimmerman SolarFlare Communications 21

22 Coding proposal Motion #2: That 10GBASE-T adopt as a 3 bits/baud/pair 4D-4Way PAM-10 code with integrated control symbols G. Zimmerman SolarFlare Communications 22

23 Power Backoff Proposal Motion #3: That 10GBASE-T adopt a powerbackoff mechanism adapted on startup for use on shorter lines levels and metrics TBD. G. Zimmerman SolarFlare Communications 23

24 Backup Slides G. Zimmerman SolarFlare Communications 24

25 Relation of Rate Loss in DFE systems under pinch-off Optimum DFE Result: 1 SNR( db) = 10*log10 exp( fbaud = 10 log10(exp(1))* fbaud B 0 ln(1 + When f_snr(f) is small for f>fbaud, increasing the baud rate does not change the value of the integral as in an AWGN channel f B 0 ln(1 + _ SNR( f )) df ) f _ SNR( f )) df ) G. Zimmerman SolarFlare Communications 25

26 ANEXT Robustness: Variability of Required Channel Capacity with Constant SNR Constraint Required Channel Capacity (Gbps) Required Channel Capacity vs Class E Channel Length for Different PAM Codes 22 2 bits/symbol: 23 db Rcv SNR 3 bits/symbol: 29 db Rcv SNR 4 bits/symbol: 35 db Rcv SNR Channel length (meters) ref: TR ANEXT = Y + 15*log10(f/100), where Y is adjusted to produce target receive SNR Channel contains 4 connectors + 10 m patch cords; length adjusted with horizontal cable span only Target SNR includes BER = 10e db coding gain 3 db margin Impairments (Class E): Echo = 55 db NEXT = 40 db FEXT = 25 db Noise = -150 dbm/hz Transmit power = 8 dbm G. Zimmerman SolarFlare Communications 26

27 ANEXT Constant ANEXT Robustness: Value of ANEXT Coupling Constant (Y) with Constant SNR Constraint Value of ANEXT Constant vs Class E Channel Length for Different PAM Codes Channel length (meters) ref: TR bits/symbol: 23 db Rcv SNR 3 bits/symbol: 29 db Rcv SNR 4 bits/symbol: 35 db Rcv SNR ANEXT = Y + 15*log10(f/100), where Y is adjusted to produce target receive SNR Channel contains 4 connectors + 10 m patch cords; length adjusted with horizontal cable span only Target SNR includes BER = 10e db coding gain 3 db margin Impairments: Echo = 55 db NEXT = 40 db FEXT = 25 db Noise = -150 dbm/hz Transmit power = 8 dbm G. Zimmerman SolarFlare Communications 27

28 Error Propagation Performance Error Prop Reduction 4D-Pairs 8st Coded PAM10, ERPX PAM10 DFE 4D-SER PAM10 DFE w/ perf FB 4D-SER PAM10 1way Vit 4D-SER PAM10 1way Vit w/ perf FB 4D-SER PAM10 4way Vit 4D-SER PAM10 4way Vit w/ perf FB 4D-SER PAM8 AWGN 4pr SER PAM10 AWGN 4D8st Vit 4pr-SER Symbol Error Rate EQ Output SNR (db), residual ISI compensated Slicer SNR (db) G. Zimmerman SolarFlare Communications 28

29 Coding Description: Encoder Transmit Data TXD n [0:11] From LFSR Sc n [0:11] TX SCRAMBLER Sd n [0] Sd n [1] Sd n [2] Sd n [3] Sd n [4] Sd n [5] Sd n [6] Sd n [7] Sd n [8] Sd n [9] Sd n [10] Sd n [11] Sd n [12] Select Point in Subset Select Subset 4D PAM10 MAPPING TXA TXB TXC TXD D D D cs n [1] cs n [1] cs n [0] G. Zimmerman SolarFlare Communications 29

30 Coding Description: Trellis Diagram Subset for each of 4 branches leaving state Convolutional Encoder Bits at time n Convolutional Encoder Bits at time n+1 Subset for each of 4 branches entering state D0 D2 D4 D D0 D2 D4 D6 D1 D3 D5 D D2 D0 D6 D4 D2 D0 D6 D D4 D6 D0 D2 D3 D1 D7 D D6 D4 D2 D0 D4 D6 D0 D D1 D3 D5 D7 D5 D7 D1 D D3 D1 D7 D5 D6 D4 D2 D D5 D7 D1 D3 D7 D5 D3 D D7 D5 D3 D1 G. Zimmerman SolarFlare Communications 30

31 Baud Rate: 55m Class E with split ANEXT model Optimum shifts towards 3 bits/baud & steepens Class E IL ANEXT Loss = 49-X*log10(f/100) (X=15, f>100, X=10, f<=100) -80 Received Residual PSDs -1.5 Uncoded DFE Margin vs. bits/baud/pair Uncoded DFEMARGIN dbm/hz signal out rem next rem alien next -180 rem fext rem echo rem awgn Uncoded Margin (db) Info bits/baud/pair G. Zimmerman SolarFlare Communications 31

32 Baud Rate: 100m Class E+ Example DFE Margin vs. info bits/baud strongly favors lower baud rates (Class E IL) -80 ANEXT Loss = 64-X*log10(f/100) (X=15, f>100, X=10, f<=100) Received Residual PSDs -1 Uncoded DFE Margin vs. bits/baud/pair dbm/hz signal out rem next rem alien next -180 rem fext rem echo rem awgn Uncoded Margin (db) Uncoded DFEMARGIN Info bits/baud/pair G. Zimmerman SolarFlare Communications 32

33 Baud Rate: 100m Class E+ Example DFE Margin vs. info bits/baud strongly favors lower baud rates (Class F IL) -80 ANEXT Loss = 62-X*log10(f/100) (X=15, f>100, X=10, f<=100) Received Residual PSDs -1 Uncoded DFE Margin vs. bits/baud/pair dbm/hz signal out rem next rem alien next -180 rem fext rem echo rem awgn Uncoded Margin (db) Info bits/baud/pair Uncoded DFEMARGIN G. Zimmerman SolarFlare Communications 33