Walter,
Thanks for your comments. Regarding the last two bullets of your summary,
* The flow in BIRD-166 does allow the Rx2 to optimize itself correctly.
* The flow in BIRD-166 give the right results.
These statements are only true when we have "all Init-Only or Dual models"
in the simulation, as you stated in your introductory sentence. As you may
recall, the discussions we had in the ATM meetings revealed that we did not
think that imposing such requirements on model makers would be feasible.
The discussions also revealed that the proposed solution in BIRD166.x makes
certain other model combinations worse, and as long as those combinations
are allowed, the proposal in BIRD166 is unable to address the conflicting
requirements for all model combinations.
This is why Ambrish ended up submitting BIRD190. It does NOT attempt to
solve the challenges we are facing, it only documents the situation in the
redriver section of the specification so that the reader would be reminded
to the issues at hand.
It seems that in order to solve these problems, we would either have to
have a "full solution" along the lines of Fangyi's proposal, or allow only
Init-only, GetWave-only or Dual models together with the proposal in
BIRD166, but the latter didn't seem to be an option based on the discussions
we had in ATM.
I would very much like to make a decision on this soon, since we are starting
to go around in circles in discussing this topic.
Thanks,
Arpad
============================================================
From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx] ;
On Behalf Of Walter Katz
Sent: Sunday, June 18, 2017 6:24 AM
To: Ambrish Varma <ambrishv@xxxxxxxxxxx>; ibis-macro@xxxxxxxxxxxxx
Subject: [ibis-macro] Re: Response to BIRD-190
Ambrish,
Got time to review my response to BIRD 190, and would like to correct what I
wrote. Changes made in line.
Walter
Ambrish,
Can you please comment on the following statements.
This discussion is limited to statistical simulations of Redriver channels with
Tx1, Rx1, Tx2, Rx2 all Init-Only or Dual models.
1. BIRD 190 states:
* The impulse response of the Rx2 model will not include all of the
effects of the upstream equalization, and therefore the Rx2 AMI_Init will not
be able to perform an accurate equalization, and therefore to ensure successful
simulations, the user should be allowed to turn off any automatic equalization
in Rx2.
2. As Fangyi so aptly demonstrated, the Rx2 equalization will include two
parts, one is a scaling (LTI) part from a CTLE and AGC, and the second an
additive part (non-LTI) from a DFE.
* If the part of the upstream equalization that is excluded in the input
to Rx2 (as specified in IBIS 6.1 and BIRD 190) is convolved with the output of
Rx2, then both the CTLE and DFE equalization will scale this part of the
upstream equalization
* If this part is not excluded from the input to the Rx2 then only the
CTLE scales the effects of the upstream equalization.
* In case 2.a. the DFE is (incorrectly) added twice
3. For those of us that find this argument unconvincing, then consider the
solution space that the user must explore to find a "best" solution if
automatic optimization is disabled as proposed in BIRD 190.
* Let us assume that the user has set all of the upstream co-efficient
to their optimum setting, and the Rx2 automatic equalization is turned off by
the user in accordance with BIRD-190.
* How many simulations would the user be required to run to find the
"best" solution.
i. Case 1
1. Assume that the Rx2 model has the following controls (typical PCIeG4,
IEEE 802.3bj)
a. 10 CTLE settings
b. 10 AGC settings
c. 4 DFE taps each having 10 tap setting
2. The answer is 10^6 simulations
ii. Case 2
1. Assume that the Rx2 model has the following controls (IEEE 802.3 400Gj)
a. 10 CTLE settings
b. 10 AGC settings
c. 14 DFE taps each having 10 tap setting
2. The answer is 10^16 simulations
* It is a requirement of our customers that our tool determine the AMI
settings that give the "best" solution. Sweeping these setting in the Rx2 as
BIRD-190 requires is unacceptable.
In conclusion:
* The flow in IBIS 6.1 does not allow the Rx2 to optimize itself correctly.
* The flow in IBIS 6.1 gives the wrong results even when Rx2 optimization
is turned off.
* The flow in BIRD-190 does not allow the Rx2 to optimize itself correctly.
* The flow in BIRD-190 give the wrong results even when Rx2 optimization is
turned off.
* The flow in BIRD-166 does allow the Rx2 to optimize itself correctly.
* The flow in BIRD-166 give the right results.
Walter
