| Thread Links | Date Links | ||||
|---|---|---|---|---|---|
| Thread Prev | Thread Next | Thread Index | Date Prev | Date Next | Date Index |
|
Dear 802.3cu Colleagues, As we will be discussing penalties today, it is worth reviewing a few of them.
TDECQ – 10log10(Ceq) The stated objective for this new spec. is to weed out transmitters with bad time domain characteristics, for example excessive pre-emphasis. The path it takes to accomplish this is a long and convoluted one. The reference RX equalizer frequency response is to some extent a representation of the inverse of the TX response. In the pre-emphasis example,
if TX is peaky, the RX equalizer will adapt to a narrower bandwidth, which results in increased TDECQ – 10log10(Ceq). Notice that for the other problem corner, the slow TX, this penalty is exactly wrong.
As pointed out in the emails below, the problem of excess TX pre-emphasis has been solved for a couple of decades through the use of upper and lower eye mask. I have enclosed the DCA eye again, showing exactly
how this is done. Unlike for TDECQ – 10log10(Ceq) which almost no one understands, this is easily understandable by everyone. It is also a direct limit of what we are trying to limit, rather than a long and tortuous detour. And this also solves the slow TX
corner problem, by limiting rise/fall time. Chromatic Dispersion (CD) Penalty The industry has known for a long time how to measure CD penalty.
ITU-T has done this explicitly for decades, although not into a ref RX, but DUT RX, which is a weakness of the ITU-T methodology (but not the CD calculation). IEEE has done this with TDP, which is a measure
of both Transmitter and Dispersion penalty. It is being advocated that TDECQ – 10log10(Ceq) is a measure of CD penalty. That is partially true. It is a measure of CD penalty to the same extent as TDECQ is a measure of CD penalty. They both measure Transmitter
plus Dispersion penalty. Calculation of TDECQ involves lot of massaging, including by Ceq. TDECQ – 10log10(Ceq) is just an additional massage step, so we could properly call it TDECQ2. Which makes it clear it is not a measure of CD. So sadly TEDECQ – 10log10(Ceq)
is not the Swiss Army knife of penalties. TDECQ – SECQ is little better estimate of CD penalty, but not a great one because SECQ is not a sensitivity measurement into ref. RX. At least it’s a step in the right direction. The Big, Missing penalty Brian Welch has pointed out multiple times the problem with TDECQ for middle wavelengths, specifically that because the CD is low, all of TDECQ is now due to the TX, which is excessive. I have raised it several
times by referencing Brian’s concerns, for example: http://www.ieee802.org/3/cu/public/Sept19/cole_3cu_01b_0919.pdf#page=5 This really is a big problem.
The Solution We already know how to do all this properly, so let’s not reinvent the wheel and just do what has worked for many years.
Now everyone is happy and we can all sing Kumbaya. Chris From: Chris Cole
Dear 802.3cu Colleagues, We will be discussing TDECQ – 10Log10(Ceq) penalty during this week’s TF meeting. It is worth resending a couple of emails I sent to the 802.3cd reflector when it was first proposed. The comments apply as
much today as when they were first made. We took a specification and measurement technique which had two decades of use and understanding, specifically upper and lower mask margin, and for no really good reason changed it into an abstract measurement,
TDECQ – 10Log10(Ceq) which few understand or can explain. Chris From: Chris Cole
Dear 802.3cd Colleagues, I received several questions last week as to how upper and lower limits of an eye mask limit overshoot. Enclosed is a 10G eye, mask and mask margins pulled off the web (apologies Pavel; it’s the first one that came up in the search). By closely squinting, notice that the the middle mask is numbered 1, upper
mask is numbered 2, and lower mask is numbered 3. The solid gray is the mask and the light gray is the mask margin. The 2 and 3 mask margin keeps the eye nearly flat. The 2 and 3 normative mask allows considerably overshoot. This can be superimposed on
the problem eye presented to the TD this week, and it’s evident how masks 2 and 3 determine the amount of overshoot. If only 2 and 3 masks are used, either a NRZ or PAM4 eye can be used. Mask 1 can be used to limit slow edges of an NRZ eye.
Mark’s expectation in his latest email is that any further changes to the cd draft will be restricted to refinements. As we start deploying 100G per lane optics, we will learn a lot more and substantive changes
are sure to come. Since volume deployment is several years away, we have the time to get it right. If we are too late for cd, we can consider other forums or future IEEE activities.
We will discuss with scope makers a couple of mask implementation questions. What are the best 2 and 3 masks for overshoot, and what’s best for limiting edges: 1) direct rise/fall time of step transition or
2) mask 1 hits of at-speed PRBS. If there is interest, we can present this in cd, otherwise in other forums.
Thank you Chris From: Chris Cole
Dear 802.3cd Colleagues, Since we adopted PAM4 modulation for 50Gb/s and 100Gb/s wavelengths, the 802.3 TFs had to develop a new set of interoperable specifications, handicapped by lack of experience with PAM4 optics. We have made
a lot of progress and the current set of specifications is converging. We have refined these specifications as we have gained experience with 50G PAM4 optics, and very recently with first 100G PAM4 optics.
We are now about to add an entirely new transmitter penalty; TDECQ - 10Log10(Ceq) ≤ TDECQ (max) (http://www.ieee802.org/3/cd/public/July18/mazzini_3cd_01d_0718.pdf#page=12)
with which the industry has zero experience, based on an entirely new abstraction, with no body of measurements, and a vague understanding of actual limitations on transmitters. TDECQ was also a novel TX penalty, the difference being that we have been living
with TDECQ for several years and have gained insight based on great deal of analysis and measurements. It is worth taking a step back to briefly review history of TX specifications. The general idea is simple. We want the transmitter eye open, with no or low errors when detected by expected range of receivers.
The approach has been two types of specifications: 1) time domain waveform limitations, and 2) BER limitations. Examples of waveform limitations are ER, noise (like Rin) and most importantly eye mask, all into a defined receiver. The BER limitation has been
expressed as a penalty against a good (ideal) transmitter into a specified receiver. In succeeding optics specifications we have reacted to the limitations of previous generations. Zero hit eye mask was replaced by stat. eye, based on the realization that sitting for many minutes (hours) in
front of a DCA, waiting for as single hit in an eye mask is not the best use of time. BTB and over fiber loopback penalty was replaced by penalty into a reference receiver realizing that tuning a transmitter and receiver to each other does not guarantee interoperability
with other vendors transmitters and receivers. Which brings us to PAM4. A traditional eye mask was found not useful because a closed eye can be opened by receiver equalizer. We attempted something new, a transmitter specification based primarily on a TX
penalty, with only auxiliary waveform limitations including ER and Rin. Over time we found this was not enough. In our lab, our system test engineer tasked with verifying 50G PAM4 optics (LR8) would periodically triumphantly announce that he found a TX waveform
that met TDECQ but broke the receiver complying with SRS. Multiple 802.3cd contributors started reporting limitations of TDECQ in guaranteeing interoperability and an understanding developed that we need to supplement TDECQ. Broadly what was required is a limit on
TX being too slow, too fast, and too noisy. In past specifications this was done by the eye mask.
Very simplistically, the inner eye mask limited excessively slow or noisy transmitters, and the upper and lower eye limits limited overshoot or excessively fast transmitters.
I proposed rise/fall time lower limits to restrict slow transmitters, and overshoot limits to restrict fast transmitters. The rise/fall time limit was adopted, but the overshoot limit was recognized as problematic.
In hindsight, an eye mask for a simplified waveform, like NRZ, or half rate should have been proposed. There was no need to reinvent the wheel, and I wish I came to this realization sooner. The eye mask is just fine, and all that is needed is a method to get
around the PAM4 eye closure. Pierce was one of those who identified the problem with TDECQ as the sole TX limitation. As part of illustrating the problem, he used the concept of TDECQ map which plotted TX penalty variants against each
other. As a result this led him to propose various complex TX penalty variants as the fix. The second TX penalty about to be adopted by the TF derives directly from viewing the problem through the prism of the TDECQ map.
While the TDECQ map is a way to look at transmitter performance, its applicability is limited. It’s not what would routinely be used to characterize or compare transmitters in the lab.
The most glaring example of why this is not broadly applicable is the example Pierce has used in support of adopting the second TX penalty spec. In reverse order, a waveform is presented which is problematic: http://www.ieee802.org/3/cd/public/July18/dawe_3cd_01b_0718.pdf#page=11 Everyone can understand this. This waveform is too peaky. It would likely have hits in the upper and lower limits of an eye mask.
This simple problem has been transformed into an abstraction, TDECQ map, which few in the TF seem to understand, at least based on the discussion in today’s TF meeting: http://www.ieee802.org/3/cd/public/July18/dawe_3cd_01b_0718.pdf#page=7
Given this reformulation of a simple problem into a complex abstraction, it is understandable why an equally complex and abstract solution is offered. If adopted, we will be left with an unprecedented standard
with two TX penalty limits. And this complex, novel approach is being done to solve a problem which has been solved in all previous optical specifications with time domain limits, like eye mask.
Before embarking on this unchartered journey we should invest a lot more effort to confirm that what has worked in the past, will not work now. One approach is to replace the rise/fall time limit with eye
mask for NRZ waveform or half rate waveform. There may be other time domain limits which work equally well. But a second TX penalty is not one of them. Thank you Chris From: Mark Nowell (mnowell) [mailto:00000b59be7040a9-dmarc-request@xxxxxxxx]
Dear colleagues, A reminder that we will start @ 8am on Tuesday First order of business will be to review the two liaison letters (now posted) and then continue the comment resolution The latest version of Matt Brown’s working document is posted for your review. http://www.ieee802.org/3/cd/public/July18/
Thanks, Mark To unsubscribe from the STDS-802-3-50G list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-50G&A=1 To unsubscribe from the STDS-802-3-100G-OPTX list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-100G-OPTX&A=1 |
Attachment:
10G DCA Eye .jpg
Description: 10G DCA Eye .jpg
