I wonder if we could use a SNR_Tx penalty based on the Tx FFE for the Rx FFE MMSE optimization?
Like:
SNR_TX = SNR_TX_0 – kappa* sum(Tx ffe samples^2)
We could use some clipping logic to set kappa to 0 for some tx FFE power level.
That seem like it would turn on the Tx FFE. It also seems that it would need some alignment to actual design.
… Rich
|
From: Hossein Shakiba <000029b7cabe11ea-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Thursday, June 26, 2025 5:36 PM
To: STDS-802-3-B400G-ELEC@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_B400G_ELEC] Regarding quantization noise - clipping probability
|
Caution: This email originated from outside the organization.
Do not click links or open attachments unless you recognize the sender and know the content is safe.
|
Hi Adee,
Thanks for sharing your experience and perspective on this topic. It is very helpful and appreciated.
In terms of criticalness of optimizing the AFE gain and ADC clipping level, I agree with you. Here are my additional two cents:
-
As I have mentioned several times, including in my interim presentation, the scope of this proposal is NOT to include the FUNCTIONALITY of quantization, but only its NOISE impact. As I am sure you would agree, modeling the functionality in
a statistical analysis platform (which is what COM is based on) is not trivial and requires a lot more study, if possible at all.
-
I fundamentally do not disagree with your suggestion (at least directionally). The exact value requires generating more data and seeing more contributions and from additional perspectives. Obviously moving in this direction increases clip
level and degrades COM as same number of bits will now quantize over a larger range of input.
-
I picked one channel case and ran a sweep of P_qc over a span of five decades to cover your suggested probability:
For this case, changing P_qc from its current value of 2*DER0 to 0.001*DER0 only makes COM by 0.13dB worse, so it is not a very sensitive function.
-
I can generate this data for all the other channels, but this will take several days. If you think this data would help I can start to lunch simulations. However, your points on the criticalness of ADC clipping level is beyond the quantization
noise and I am not sure if the data can help. So, if this is the direction you would want to go, then the clipping level probability call must be made independent of quantization noise. The only constraint, would be not to exceed the upper limit of 2*DER0,
which is obvious.
-
In light of the discussion and straw poll today, and hopefully with the help of my comments in this email, I suggest we continue with the comment against D2.0 with P_qc = 2*DER0 (as it is now and was during the poll) and look into changing
it after new contributions are presented later. I am more than happy to cooperate with you and others towards a final resolution.
Regards,
Hossein
From: Adee Ran (aran) [mailto:0000147b29386f6c-dmarc-request@xxxxxxxxxxxxxxxxx]
Sent: June 26, 2025 12:38 PM
To: STDS-802-3-B400G-ELEC@xxxxxxxxxxxxxxxxx
Subject: [802.3_B400G_ELEC] Regarding quantization noise - clipping probability
In Hossain’s presentation today, it was mentioned that the clipping probability was set to about the target DER. I did not comment on that due to the expected length of the other Q&A.
In my experience, ADC clipping is a source of large errors that have a high probability of creating errors. Moreover, clipping events are often related to low-frequency effects such as baseline wander and/or long runs of outer symbols.
Thus the errors created are correlated and can severely impact the FEC performance.
As a result, in practice, gain control loops for ADC drive to have much lower probability of clipping than the allowed BER; I’ve never seen anything higher than 1e-6 in FEC-aware receivers, and even 1e-9 is common.
I think the high clipping probability also allows signals with large peak-to-average ratio at the ADC input, which may explain why Tx equalization is not used. When Tx equalization is used together with the CTLE, the distribution would
be more bounded and allow for much lower clipping probability for a given quantization noise.
The straw polls did not include the detail of clipping probability; We need to adopt a value.
I would recommend starting with a moderately conservative assumption of 3 orders of magnitude below DER0. Further analysis would be required and welcome.
</Adee>
To unsubscribe from the STDS-802-3-B400G-ELEC list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G-ELEC&A=1
To unsubscribe from the STDS-802-3-B400G-ELEC list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G-ELEC&A=1
â??
To unsubscribe from the STDS-802-3-B400G-ELEC list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-B400G-ELEC&A=1