| Thread Links | Date Links | ||||
|---|---|---|---|---|---|
| Thread Prev | Thread Next | Thread Index | Date Prev | Date Next | Date Index |
|
Hi Brian, Tingting, (All), Chiming in. Full disclosure – I have no preference for PAM2 vs PAM3 training.
Just some opinions here based on some statements made. We can all agree or disagree.
1) Finding symbol/frame boundaries is slightly simpler using PAM2, but once you work out the scheme using PAM3
the time to lock and gate count difference is negligible. If you are clever you can do the symbol and frame boundary
lock at the same time. The time to find the training frame boundary is the thing that dominates the lock time. So in my opinion PAM2 vs PAM3 is not an issue here.
2) Intuitively it is always easier to train with PAM2 vs PAM3. The comment about the complexity switching from PAM2
to PAM3 I do not believe is true. It looks like we are close to consensus on the training frame and info field.
With the contemplated training sequence, we know exactly which symbol PAM3 will start since there is a countdown
in the training. We need to know this boundary anyway since we need to know the exact symbol when to stop interpreting
training frames and start interpreting data frames. So in my opinion there is no complexity difference between
transitioning from PAM2 to PAM3 vs staying in PAM3 all the time.
3) The question about the spectrum between PAM2 and PAM3 being too different I do not agree with without seeing
some data. There was a 4B6B that people say is too few symbols. We can compose a 5B6B code that can limit running disparity
and get 32 codes instead of 16. Intuitively, I feel that end of the day the spectrum would be similar enough.
This was never a problem with 1000BASE-T that trains on PAM3 and has data on PAM5. Same for 1000BASE-T1 that
trains on PAM2 and had data on PAM3. So I reserve judgement on this. I think the burden is on Tingting to show the
spectrum of a 4B6B or a 5B6B is similar enough to 8B6T.
4) Assuming the spectrums between PAM3 and PAM2 training sequence is similar enough and don’t have any material emissions differences,
I think the real crux of the debate here is whether training on PAM2 reduces the probability of a retry vs PAM3 justifies the
additional complexity of additional circuitry to do PAM2. Note that the additional circuit here in terms of gate count is very
small mostly to do the 4B6B or 5B6B mapping that otherwise would not be needed. Everything else would be the same
(i.e. the scrambling, the info field, the countdown, the state machines). Since the timeout on link up is 100ms, it means
that every retry on startup incurs 100ms penalty.
I don’t have an answer here whether the retry issue is material or not especially
at 500 meters to be something we worry about. It would be good if people can bring in data to show if retry is really a problem
or not using PAM3 vs PAM2 especially over challenging channels.
The arguments about cost, complexity, startup times (other than retry) should not be an issue here in my opinion.
Thanks, William From: Murray, Brian <Brian.Murray@xxxxxxxxxx>
Tingting Please find my answers to your questions below. > impacting on the 6T boundary decision. In the worst case, it wastes 6x synchronization time. The symbol alignment is very fast and is a tiny portion of the start-up time, much, much less than 1%. So I don't see this as a significant factor. > If PAM training using PAM3 blind equalization while data mode uses PAM5 partial response, will this work ok or not? Yes, it does work. Blind equalization is very commonly used for PAM-3 modulation in multiple IEEE standards and has the advantage of simplicity. > Will PAM2 make this simpler? 4B6B may be not the best coding, but I do not think additional coding for PAM2 will increase hardware complexity significantly. I am not concerned about the hardware complexity, which I agree will be a small addition. What I am concerned about, and what I believe most people who raised the same question for your presentation in Hamburg, is data correlation with PAM-2 training. Given that it is mandatory to support
running disparity, there are only 48 possible PAM-2 codes (out of 256). Your gave an example of using the 16 codes which have a very different frequency spectrum from the 256 codes used during data. Even using 32 codes, there is still a very significant problem
with data correlation. So we then need a scheme (and start-up time) to later switch from PAM-2 to PAM-3 and we have the risk that the equalizer and each cancellers are not correctly trained when we change to PAM-3. Furthermore strong correlation in the PAM-2
sequence will result in spurs in the power spectrum. This is highly undesirable and runs the risk of creating EMC issues. Late changes were made in the 10BASE-T1L standard to eliminate correlation effects far less severe than would arise when using a PAM-2
sequence. We had given serious consideration to the possibility of using PAM-2 but concluded that the complication of figuring out how to switch back to PAM-3 and verify that the equalizer and each cancellers
are not correctly trained would negate any benefit and overall we would be worse off. > Less link up time is beneficial for the application. First of all I don't believe PAM-2 will result in lower link up time and may result in longer linkup time or links that startup and then fail, which would be worse. Secondly, lower link time is only an advantage if the application can benefit from it. We are proposing to target 100ms including Auto-Negotiation. This is the same as 100BASE-T1 and 1000BASE-T1.
The system needs to be designed around the normal maximum time to link which is comprised of the ANeg time plus the link_fail_inhibit time. A lower time to link is not an advantage as the system needs to be designed for the worst case. So unless we plan to
further reduce the link_fail_inhibit time I don’t really see that link time is a factor that needs to be considered. Brian From: zhangtingting (O) <zhangtingting59@xxxxxxxxxx>
Hi Brian, I wish we can achieve consensus on PMA training for the residual details so that we can make D1.0 possible in November. Regarding reusing 8B6T PAM3 for training mode,
I have the following concern: 1. The advantage of simplifying hardware is obvious, but the training efficiency is low, considering the higher error impact on scrambler synchronization,
consequently impacting on the 6T boundary decision. In the worst case, it wastes 6x synchronization time. 2. If PAM training using PAM3 blind equalization while data mode uses PAM5 partial response, will this work ok or not? 8B6T with partial response
is a good idea to achieve coding gain without FEC. We should think about a reliable training scheme with 8B6T partial response taken into account. Otherwise, no benefit comparing it with the 8B6T for 100BASE-T4. Will PAM2 make this simpler? 4B6B may be not the best coding, but I do not think additional coding for PAM2 will increase hardware complexity significantly. Less link
up time is beneficial for the application. Hear from you. 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 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 |