Hi Chris,
I agree for power savings.
Also once we start to talk about CPO/NPO approach (Straw poll #2), we need to be mindful of which channel we are talking
about in terms of saving pJ/bit because CPO/NPO goes out of the box?.
Is it ~20 inches of channel or 50 meters of channel
BR,
Joshua
From: Chris Cole <chris@xxxxxxxxxxxxxxx>
Sent: Wednesday, September 28, 2022 11:28 AM
To: STDS-802-3-B400G@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_B400G] Oct 2022 Series Webpage Update
Hi Brian
I disagree. Bypassing the FEC in the module is not the biggest power savings.
The biggest power savings comes from 20dB SerDes in both the ASIC and the module, vs. a 36dB SerDes.
Chris/Ali,
I think the biggest power savings in the module is going to come from bypassing the FEC. As such I think there is value in trying to push the medium loss case
as high as you can while still getting the needed DER improvements, as it leads to more modules in any given system that could run in bypass mode and hence more power savings.
I am mindful of that fact that in both cases the EQ requirement will likely be quite heavy, and it will be interesting to see how much that could scale for the
“CPO/NPO” reach.
Thanks,
Brian
Hi Upen,
Your numbers are too high. They reflect legacy design practices and would inhibit innovation and limit meaningful performance gains like power reduction.
I think Ali, proposed loss range is too low for High radix designs including CPO/NPO
I would adjust those ranges to
-
36-38 dB for large radix
-
25-30 dB for medium radix and other anticipated ( including CPC/NPC)
-
18-25 db for CPO/NPO
-Upen
Hello Adee:
We have at least 3 classes of AUIs
I. Conventional package PCB up to 36 dB
II. Advance package with cable or CPC up to 22 dB
II. Co-packaged with or without 1st level package ~18 dB or ~12 dB
The difference between 1 and 2 is certainly greater than just 1 order of magnitude in DER. AUI type 1 would require something like 40 UI span equalizer, possibly
MLSE turned on, and FEC termination.
AUI type 2 likely can operate with an equalizer with 12-16 UI span and DER of ~1E-5. AUI type 3 exact EQ will be dependent if we have the 1st level package or
not, so there is a range of possibility but overall simpler than AUI type 2.
AUI Type I, II, and III are different specifications and each should have their own clause, but some product in the market place be a superset that can support
more than one AUI types.
Following today’s call and straw polls, I’m thinking of doing some analysis for the “lower loss” AUI.
I can see the two specifications going in a few directions:
-
“Long AUI”: up to 36 dB with termination of the RS544 (DER≈1e-4); “Short AUI”: up to x dB without termination (DER≈1e-5). Essentially segmented/concatenated, but with the same bits on the optics for both. Electrical specs may be the same,
except for the BER/DER parameter. This would enable modules with no RS-FEC implementation – lower area, power, AND latency.
-
-
“Long AUI”: same as above; “Short AUI”: up to x dB with weaker SerDes assumptions (equalization etc.) but same DER target. Both assume termination of the RS-FEC. Different electrical specs. “low loss” modules can have lower power and area,
but a relatively small effect on latency.
-
A combination of the two above – if something can be gained (I’m not sure)
-
Something else?
Is there a preferred direction?
Yes, if you have a wide loss dynamic range then of course calibration or training is critical. But the more important question is why have it in the first place.
For many generations of copper interconnect, the loss was dominated by one segment of the channel. In that case, one large gain stage works just fine. But this is no
longer the case. The packages at both ends, the PCB trace(s), the cable if there is one, all are significant contributors.
Basic math tells us that a single large gain stage is a power inefficient solution to this problem. That's nothing new. Nearly two hundred years ago, people figured
that when stringing telegraph lines they needed to intersperse them with repeaters. They didn't have the math to help them understand why. However, we do, so what's our excuse?
A DAC at 224G/lane is not the lowest power solution; it's the highest power solution. To add insult to injury you and Gary want to force this overgrown gain on C2M channels
that don't need it. Training or calibration hardly makes it palatable.
I don’t need to have a self-driving car to get anti-lock brakes.
The challenge here is calibration of control. A scheme that is robust to miscalibration is enabling.
Juniper Business Use Only
[External Email. Be cautious of content]
That's a great idea. Swiss Army knife do-it-all engineering solutions have always resulted in lowest cost and power.
In the same vein, we should adopt Coherent ZR for all optical PMD reaches. After all, it's powerful link training is fully capable of adapting to the worst case
multi-hundred km channels, and all shorter applications down to tens of meters.
I indeed hope it becomes mute.
The discussion here was recalling that AUI was devised with two loss targets for 100G electrical lanes. We avoided detailed parameter programming. To embrace
even more complex configuration, then link training indeed becomes attractive.
Juniper Business Use Only
[External Email. Be cautious of content]
Does all of this become mute when we move to electrical link training for 200G C2M ?
I thought the intent of link training is that rather than having a pre-defined set of parameters based on the worst-case channel (or in the case of 3ck the two
worst channels of AUI-S and AUI-L), that the parameters are adapted to match the actual channel ?
I agree deeper thinking needed. How we indicate and configure AUI support is an important problem too or interop between host and module may never happen despite
all the excellent deep thinking.
Juniper Business Use Only
[External Email. Be cautious of content]
I am glad we have zeroed in on the critical aspect of what needs to be done.
I was trying to distract us with side topics like the loss budget and applications scenarios, but perhaps that will just fall out of getting the correct host
id.
AUI-S and AUI-L are not PMDs in SFF-8024. They are Host Electrical Interface IDs. At present, they are seen as two configurations of one implementation. Please
see IEEE 802.3ck, where they are defined as such.
How do you see mapping “classes of AUIs” to Host Electrical Interface IDs?
Juniper Business Use Only
[External Email. Be cautious of content]
AUI-S and AUI-L is are two settings for the same retime interface (4/5 taps FFE with RX CTLE+4DFE taps), assuming we will introduce some level of adaptive transmitter
via CMIS-LT/CL136 you could have 1000’s of valid TX settings.
It is unfortunate that in SFF-8024 AUI-S/L are defined as if they were different PMDs, they are just two settings for one PMD!
AUI-I, II, III, and IV all are in my opinion within the scope of the 802.3df.
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
I thought we already agreed to leverage IEEE 802.3ck. I was simply pointing out that we have for instance 100GAUI-1-S and 100GAUI-1-L from IEEE 802.3ck, which
I understand we will similarly develop 800GAUI-8-S and 800GAUI-8-L.
Juniper Business Use Only
[External Email. Be cautious of content]
At this point for 200g lanes - no decisions have been made. Perhaps there is progress elsewhere but that does not apply within our project.
As always we are a contribution driven organization.
John
When we finish the first draft we will send it to you.
We already have AUI-S and AUI-L for two ranges essentially of loss, not just one. Regardless, we do have more kinds of implementations needing support from the
standard.
Non-Juniper
[External Email. Be cautious of content]
Dear IEEE 802.3df Participants,
I trust everyone rushed to review next week's material as soon as John announced its availability, and discovered the excellent exposition by Mr. Lusted. For
those with limited time and trying to decide which presentation to read, may I kindly call your attention to:
As always, Kent is gently steering us towards wisdom, which for some of us who prefer the more direct approach, is a bit too old school.
What's clear is that a single large loss AUI C2M is no longer sufficient and we should write multiple specifications to adequately meet the need of mushrooming
applications, specifically:
-
Traditional large loss for LR backplane, CR passive DAC, and VSR front pluggable
-
New XSR for NPO, twinax-over-PCB, active copper, and XSR front pluggable
Obviously the exact loss is TBD, however a good start for the XSR value are the shorter reach examples in last year's presentation by Sam and Nathan.
In off-line discussions, there has been a lot of interest in XSR's potential to save power, which we will direct into a proposal for the next meeting. For those
that would like to join us in contributing or reviewing, please send me an email so that we can put you on copy as we iterate a draft.
The proposed PAR modification to IEEE P802.3df and the proposed IEEE P02.3dj PAR have been uploaded.
For the proposed PAR modification to P802.3df, the date for Item 4.2 could not be entered as presented
to the Task Force. The date proposed to the Task Force was Nov. 2023, however, the MyProject system will not permit entry of a date earlier than Jan 2024. Mr. Law has contacted IEEE SA Solutions Support regarding this issue. At this time the date of Jan
2024 has been entered for Item 4.2 with the following noted entered for 8.1 –
Item 4.2: The earliest date that Myproject would allow to be entered is Jan 2024. This prevented the
entry of the desired date of Nov 2023. (A bug report has been submitted to IEEE SA Solutions Support. This item of the PAR will be updated to the desired date of Nov 2023 and this note deleted immediately upon a solution being provided.)
For the consideration of the Task Force - the motion to adopt this proposed PAR will be modified to
allow modification of the PAR in accordance with the note above.
Chair, IEEE P802.3df Task Force
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
To unsubscribe from the STDS-802-3-B400G list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G&A=1
|