On Figure of Merit in PAM4 Optical Transmitter Evaluation, Particularly TDECQ

Transcription

1 On Figure of Merit in PAM4 Optical Transmitter Evaluation, Particularly TDECQ Pavel Zivny, Tektronix V1.0

2 On Figure of Merit in PAM4 Optical Transmitter Evaluation, Particularly TDECQ A brief presentation assessing current state of the art in the evaluation of optical transmitters for PAM2 and PAM4 links. Focus on transition from TDP and Mask Test to TDEC and TDECQ, some comments on alternatives. Practical outcomes from 100G and 400G systems with 28GBd and 56GBd signaling in optical Ethernet, 112 GBd outlook 2

3 How did we get to TDECQ: Mask test Test such as Mask test, TDECQ, etc. are of little interest to the end user; the end-user only cares about BER For a number of reasons that s rarely directly measured in production. A quality-of-transmitter measurement tool developed in the 1950 s was an eye mask (made with a wax pencil on a CRT front-plate). (Bell Labs idea?) When real-time digitizing enables standard mask-test, a completely automated method of mask test becomes possible. Unfortunately the statistical nature of an high waveform-rate acquiring CRT disappears it takes a while for everyone to realize that the new, PkPk nature of the mask test is not better than the relative intensity of a wax-pencil and an analog screen This digital mask test has a notoriously unclear relationship to the BER. Likely was better in the original test on analog oscilloscope. 3

4 From: mask test to TDEC to TDECQ: Mask test with Hit Ratio One improvement to the Mask test is the hit-ratio; comes close to the original analog. Mask testing used to be a 0-hit allowed Current standards allow a small fraction of hits around 1E-5 Often there also is a minimum number of hits recommended or required So 1. Why is this ratio a good idea? 2. Why you should NOT test to 0 hits 3. What happens when you test with more hits or if you test with less hits? 4

5 Mask test without and with hit ratio The Orange curve: IDEAL (perfectly separates good quality devices from bad quality devices Blue: traditional margin test: only POORLY separates bad from good. Many FAILED good, many PASSING bad. Green: hit ratio mask test. Better than Blue, but not as good as Ideal. Note: the green line shifts relative to the blue line. That means that you need a different mask for ratio mask than for old-style mask (to get the same test) 5

6 Mask test: you should not test to 0 hits Because: It is a worse tool for separating good from bad Also: More waveforms à more fails. That s a bad test. The test you want is: More waveforms àhigher accuracy so use the Mask test with hit ratio. Those are the two reasons why you should NOT test to 0 hits. 6

7 Recap: 1 of 2: TDECQ: a measurement methodology for PAM4 optical signals TDECQ measures the merit of an optical transmitter DUT as follows: For a given TX DUT, imagine a perfect TX with the same OMA outer at the RX (i.e. after the channel) as the TX DUT The TDECQ is a ratio of: - certain vertical eye opening measure of an (simulated) perfect TX to - the same vertical eye opening measure of an physical DUT TX as it performs through its worst case optical channel (max. allowed reflection at the worst polarization, worst dispersion allowed) 7

8 Mask test final thoughts Mask test still doesn t correlate well to the user s figure of merit (BER) Mask Margin not related to link performance It doesn t accommodate equalization well so it figures that it would be replaced! 100GBASE-SR4 defines TDEC PAM4 standards ( 400G ) define TDECQ Reason to replace mask test: Its correlation to BER (bit error rate) (or for PAM4 to SER, symbol error rate) is not direct enough Margin is not in terms usable to the link designer 8

9 100GBASE-SR4 and TDEC (TDEC: Transmitter and Dispersion Eye Closure penalty) TDEC replacing both mask test and TDP, Transmitter and Dispersion Penalty (TDEC: Transmitter and Dispersion Eye Closure penalty) Some comparisons to 10GBASE-LRM, an early equalized optical standards TDEC replacing both mask test and TDP, Transmitter and Dispersion Penalty Figure: IEEE 802.3bm 9

10 TDEC compares the vertical bathtub height between your DUT and an ideal transmitter. Relatively complicated processing of four histograms to find the eye closure to BER of interest This accumulates the histograms (PDFs, probability density functions) into a CDF, cumulative density function) equivalent to a BER contour, not to a waveform density 10

11 TDECQ measurement OMA,outer= amplitude of PAM4 signal Smaller TDECQ is better (penalty). More acceptable receiver noise is better. 11

12 TDECQ measures the eye opening by adding noise till closure same for TDECQ Note on the vertical eye opening measure: how much added Gaussian noise can be added to close the eye to a given symbol error rate. Since the ideal TX is simulated, it s not really physically measured in the method, just accounted for. Removal of the (difficult to obtain) ideal TX is a big advantage over the predecessor method (TDP, Transmitter Dispersion Penalty). Since the ideal TX noise is just accounted for, and since the vertical eye opening measure is how much noise can be added, the ratio reduces to an expression with noise in the denominator ( R ) plus constants and spectra width related values. The oscilloscope receiver includes a 5-tap, T-spaced FFE optimized for minimum mean square error. (7 taps for 100Gb/s lanes) 12

13 TDECQ measurement setup Polarization rotator PMD Tx Optical splitter Test fiber O/E CRU Oscilloscope with B-T f resp. and pattertn trigger Variable reflector Reference Equalizer TDECQ algorithm DUT Optical channel Optical oscilloscope with SW After 802.3bs D2.1 Figure TDECQ conformance test block diagram 13

14 Reference Equalizer Optimizes TDECQ result Before EQ After EQ 14

15 Current problems TDECQ measurement is ok, but The Equalization is an issue. See previous slide the equalization to a large degree compensates for dispersion. With some license it might be ok to say that TDECQ w/o equalization is ok for system that don t suffer large dispersion (narrow lasers, and/or short links) Currently Tek can disable the IEEE equalization. This might be a path to a subset of the 802.3bs test if one is needed. 15

16 Concerns re TDECQ: TDECQ measures TDECQ. What do we optimize? The calculation of TDECQ requires a set of FFE tap coefficients. The method for determining the taps is to minimize the mean square error (MMSE) of the equalized waveform. Unfortunately, sets of taps with relatively small differences in MSE can have relatively large differences in TDECQ. Also, the MMSE optimized taps do not necessarily produce the minimum TDECQ. Both of these problems causes a repeatability issues with TDECQ. Having found many sets of optimized taps, where each set was optimized using MMSE, the next task is to find the overall optimal set of taps. The most obvious method would be to pick the set with the smallest MSE. Plotting the TDECQ and MSE for many sets of taps shows (see next page) that small changes in MSE can produce large changes in TDECQ and that the minimal MSE point doesn t always correspond to the minimal TDECQ. This sensitivity and the difference in minimal points is one of the reasons for the repeatability issues seen in the TDECQ measurements. 16

17 . TDECQ optimized MMSE, but it measures TDECQ - this has been changed Here is an example of a partially convergent MMSE TDECQ; note that several near-optimal MMSE solutions yield rather disparate TDECQ results Again this is a penalty, so smaller is better TDECQ, MSEE sample MSEE and TDECQ correlation near the minimal MSEE TDECQ experiment # MSEE Figure 1: MSE, TDECQ with different phase and cursor positions 17

18 TDECQ measurement vs oscilloscope BW Tek s result are similar as experiments done on 56 GBd transmitter at Cisco. Here showing the Cisco results, as per IEEE mazzini_01_0317_smf; by Marco Mazzini. (trace Highlight by prz) All TDECQ plots show similar trends by varying RX BW. Longer equalizers than reference (> 5 T- spaced, > 7 T/2) have < 0.5dB delta TDECQ. NOTE: sufficient oscilloscope BW for 26 GBd has been available (e.g. Tek 80C15). BW for 53 GBd: only solution: use Tek 80C10C ultra-fast module 18

19 Measurement bandwidth going from 0.75 *BW_el to 0.5 * BW_el Note the difference between 25 Gb/s NRZ lanes and the 25 GBd PAM3 lanes NRZ Pre-Emphasis Equalizer PAM4 D/A CR Nyquist Low Pass A/D FFE 1-T 19

20 The end Summary: TDECQ is here to stay. It might be less intuitive than the Mask test but it s sufficiently similar and it s an improvement for the whole process. Thank you any questions? 20