| Thread Links | Date Links | ||||
|---|---|---|---|---|---|
| Thread Prev | Thread Next | Thread Index | Date Prev | Date Next | Date Index |
|
Tingting
Apologies for not responding sooner, but I was finishing my presentation for next week’s meeting.
On point 1, I believe you are wrong. The complexity you have described is a full search of the codebook of 430 6-tuples for each received 6-tuple symbols. The whole purpose of Maximum Likelihood and the host of papers going back over 50 years to Viterbi and
many others is efficient algorithms to achieve the same or very close to the same gain as an exhaustive full search using a Euclidean distance. I do not believe any PHY has approached the solution using an exhaustive codebook search. I expect the complexity
of the ML detector to be much, much less than the echo canceller.
On point 3 I don’t agree. The primary purpose of using 8b6T is to mitigate error propagation as error propagation will greatly degrade the performance of an RS FEC. It also has the advantage of increase distance between the partial response sequences and has
an added advantage of that this can be used to improve BER. 8b6T is not over designed for the process industry and in fact I believe is necessary for the process industry as that application probably requires the longest reach and also requires low latency.
The improvement in BER that can be achieved with 8b6T is necessary to get to 500m at better than 10-10 BER on the 802.3dg IL model. Your simulation results from May (19.6 dB) and my simulation results from June (19.0 dB) have shown that 4b3T coding does not
achieve the required SNR margin at 500m. Hence, the need for 8b6T with partial response.
On point 2 I do not agree with your conclusion. Yes, for an ideal channel with only AWGN, simulated BER should be close to the theory. In fact for an ideal 1 + D channel, AWGN and an exhaustive search of all possible PR sequences to select the best (most likely)
sequence it should be very easy to show that the Symbol Error Rate is lower searching the full codebook compared to a Quinary Hard Decision and a lookup table of the inverse mapping OR compared to recovering the normal DFE output with a single tap DFE with
the tap = 1 and a using Ternary Hard Decision. And in all cases the MSE and SNR relative to a fixed transmit symbol power is the same.
I was confused by your statement in the presentation on the email reflector (slide 3) the minimum distance between the last 5 symbols of the 6 tuple PAM 5 is 1 not root(2). Therefore, no performance gain. This seemed to imply that you are only using 5 symbols
in the detection. But I see you have clarified in your email answer to George that you are using all 6 symbols to determine the most likely sequence. And I see in your presentation this week, that you state that each received 6 tuple should be used in the
ML detection, in order to increase the minimum distance root(2).
So given that the distance looking at the PR sequences over 6-tuples is greater than just looking just 1T at a time, the BER will be significantly lower for the same AWGN. Even though the MSE is the same in each case and thus the SNR relative to a fixed transmit
symbol power is the same in each case. And this is what our simulations show.
I look forward to continuing the discussion next week.
Brian
From: zhangtingting (O) <00001e92abc91102-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Tuesday, July 9, 2024 15:30 To: STDS-802-3-SPEP2P@xxxxxxxxxxxxxxxxx <STDS-802-3-SPEP2P@xxxxxxxxxxxxxxxxx> Subject: [802.3_SPEP2P] 答复: DATE CHANGE for upcoming 802.3dg meeting at the July 802 plenary
Hi Brian,
The presentations for June Ad Hoc show kind of consensus on RS-FEC, dual-mode PCS, and block coding. The main difference is using 8B6T or 4B3T. Regarding 8B6T with partial response, I have the following comments:
1. ML detection is used for 6-tuple decision to get the performance gain. For each received 6-tuple with noise, every Euclidian distance calculation needs 6 multiplication. Considering 430 reference 6-tuples, 2580 multiplications are needed. The detection complexity is even 2x higher than DEC. Please correct me if I am wrong.
2. 8B6T performance gain is achieved by referring to a calculated theoretical SNR of 17.9dB. For an ideal channel with only AWGN, simulated BER should be close to theory. However, the results show that in the case of AWGN, 8B6T with partial response requires higher SNR than 4B3T at the same BER. More details of the simulation can be found from the attachment.
3. The primary reason of using 8B6T is to achieve coding gain for long-reach transmission. Considering FEC for long distance and high ML detection complexity, 8B6T seems to be a bit overdesigned for 802.3dg. After all, process industry requires a low-power 100BASE-T1L PHY.
Thank you very much.
Best wishes, Tingting
To unsubscribe from the STDS-802-3-SPEP2P list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-SPEP2P&A=1 |