Re: [802.3_50G] review of baveja_3cd_01

Thread Links	Date Links
Thread Prev	Thread Next	Thread Index	Date Prev	Date Next	Date Index

Hi Jonathan,

Thank you for the analysis method. I will perform the correlation study using the algorithm you described.

While I would love to mention how many FFE taps are there in the PAM-4 chip, I do not know the answer to that question. Even if I did, it is not my IP so it is not fair for me to comment. What I can do is if there is another analog or digital DSP with less than 9 taps, I would be happy to have the devices measured using the same setup but using different DSP as a side by side comparison.

If anyone can provide such a DSP eval board, under NDA , we will be happy to do the work and do head to head comparison with FFE taps > 9 and FFE taps less than 9 .

Thanks and BR,

Prashant P Baveja, Ph.D

Deputy Manager, R&D

Applied Optoelectronics, Inc. (NASDAQ: AAOI)

13139 Jess Pirtle Blvd
Sugar Land, TX 77478
USA
www.ao-inc.com

+1-281-295-1800 Ext. 287
+1-281-295-1888 (fax)

Prashant_Baveja@xxxxxxxxxx

Copyright 2016, Applied Optoelectronics, Inc.
This electronic mail transmission (together with any attached documents) is strictly confidential and intended solely for the addressee (including addressee's employing organization, assigns, and affiliates). It may contain information that is proprietary to Applied Optoelectronics, Inc. If you are not the intended addressee nor associated with the intended addressee's organization, you must not use, disclose or copy this transmission, and are asked to notify the sender of its receipt. Please be further advised that the unauthorized interception or retrieval of e-mail may be a criminal violation of the Electronic Communications Privacy Act. The sender disavows any intention to create an agreement or an electronic signature by means of this transmission.

From: Jonathan King [mailto:jonathan.king@xxxxxxxxxxx]
Sent: Tuesday, November 07, 2017 3:28 PM
To: Prashant Baveja; STDS-802-3-50G@xxxxxxxxxxxxxxxxx
Subject: RE: review of baveja_3cd_01_1117

Hi Prashant,

Thanks for confirming the feedback. I do appreciate how much work went into gathering the data and presenting it – thank you!

For data analysis, when trying to determine a correlation value on noisy data with a truncated range (ie, where the scatter is a significant fraction of the range) the simple linear fit in excel is generally going to be very misleading.

I prefer to determine RMS variation a hypothetical curve, for example I would hope that TDECQ is a good predictor of receiver sensitivity (for a receiver with the same EQ as the ref EQ) so it would have a 1:1 slope with measured sensitivity.

My posted review plotted a 1:1 slope graph (y=x + c) for which ‘c’ was optimized to minimize the RMS error for all the data points in each graph. This RMS value gives you an indication of how good a predictor TDECQ is for Rx sensitivity.

It sounds like a discussion about what the minimum number of taps should be for real EQ implementations would be useful.

I think it’s also very important (for credibility) to state what functionality was used when making the receiver sensitivity measurements – for example 12 tap T spaced FFE + 2 DFE.

(But use of DFE’s would definitely help low bandwidth parts a lot, but lead to error propagation, so comparing results at 2.4e-4 would no longer be valid).

Again, I appreciate how much work went into gathering the data, and presenting it – thank you!

Best wishes

jonathan

From: Prashant Baveja [mailto:Prashant_Baveja@xxxxxxxxxx]
Sent: Tuesday, November 07, 2017 12:56 PM
To: Jonathan King <jonathan.king@xxxxxxxxxxx>; STDS-802-3-50G@xxxxxxxxxxxxxxxxx
Subject: [EXTERNAL]: RE: review of baveja_3cd_01_1117

Thank you for the review.

The trouble is that the data point looks as an outlier as the data is sparse. Doing the measurement on 10 devices was significantly time consuming. That device happened to have best RX sensitivity. Upon further review of chip level data, that device is part of normal distribution but has lower bandwidth than its peers.

The feedback I have is

Measure more devices, lets call it 20, reflecting normal wafer level distribution, this will allow more data loading on low bandwidth side
Turn on the TDECQ optimizer
Measure with and without time center optimization.
Rereport the correlation or lack of thereof at next ad-hoc