IBIS Macromodel Task Group

Meeting date: 15 February 2022

Members (asterisk for those attending):
Achronix Semiconductor:     * Hansel Dsilva
Amazon:                       John Yan
ANSYS:                      * Curtis Clark
                            * Wei-hsing Huang
Cadence Design Systems:     * Ambrish Varma
                            * Jared James
Google:                       Zhiping Yang
Intel:                      * Michael Mirmak
                              Kinger Cai
                              Alaeddin Aydiner
Keysight Technologies:      * Fangyi Rao
                              Majid Ahadi Dolatsara
                              Ming Yan
                              Radek Biernacki
                            * Rui Yang
Luminous Computing            David Banas
Marvell                       Steve Parker
Mathworks (SiSoft):         * Walter Katz
                              Mike LaBonte
Micron Technology:          * Randy Wolff
                            * Justin Butterfield
Missouri S&T                  Chulsoon Hwang
Siemens EDA (Mentor):       * Arpad Muranyi
Teraspeed Labs:             * Bob Ross
Zuken USA:                  * Lance Wang

The meeting was led by Arpad Muranyi.  Curtis Clark took the minutes.

--------------------------------------------------------------------------------
Opens:

- None.

-------------
Review of ARs:

- Arpad to submit BIRD211.4 to the Open Forum.
  - Done.
  
- Walter to send out BIRD213.1 draft 12 incorporating the meeting's changes.
  - Done

--------------------------
Call for patent disclosure:

- None.

-------------------------
Review of Meeting Minutes:

Arpad asked for any comments or corrections to the minutes of the February 8th
meeting.  Ambrish moved to approve the minutes.  Randy seconded the motion.
There were no objections.

-------------
New Discussion:

Specifying FEXT and NEXT in the impulse_matrix for statistical flow:
Hansel shared a slide detailing his question/proposal.  It noted that in IEEE
802.3ck, the COM tool only applies TXLE to the through-channel and FEXT
responses.  It does not apply TXLE to NEXT.  Hansel observed the IBIS currently
does not categorize the types of crosstalk in the crosstalk rows of the
impulse_matrix parameter of AMI_Init.  He asked whether we should change the
specification to state that the rows of the impulse_matrix contain the
through-channel response first, followed by FEXT responses, and then followed
by NEXT responses.

Wei-hsing said this would also require some way to indicate how many rows were
FEXT and therefore how many were NEXT.  Fangyi observed that this proposal would
also interact with the newly submitted BIRD211.4, AMI Reference Flow
Improvements.

Walter said that the fact that the 802.3ck COM specifies that Tx equalization
only applies to FEXT has nothing to do with NEXT and FEXT in our world.  He said
in the 802.3 standards the definition of FEXT was narrower, and it referred to
aggressor and victim signals traveling from identical transmitters on the same
chip to identical receivers on a single chip.  Everything else is classified as
NEXT.  He said this has nothing to do with what an Rx must do when it's getting
crosstalk contributions from all sources.  Therefore, the AMI Rx must apply its
equalization equally to all contributions.

Walter said if an EDA tool were to want to employ special processing based on
the IEEE standards' definition of FEXT and NEXT, it was free to exclude Tx
equalization from transmitters on NEXT channels, likely because it didn't know
the settings.  FEXT aggressor channels, on the other hand, have identical
settings to the primary channel.  Walter said COM excluding Tx equalization for
the NEXT terms was simply a case of COM wanting to be conservative.  FEXT terms
can be handled specially because of the specific definition of FEXT, and for all
the NEXT terms the conservative approach is to apply no equalization to the NEXT
channels.

Arpad asked why we would care about this if it's specific to COM.  Michael asked
if there's a scenario where the IBIS 7.1 interpretation of the impulse_matrix is
a problem.  Fangyi asked the rhetorical question, "In the physical device, how
could the Tx apply a certain equalization to just one component of the cross
talk and not the other?"  Walter agreed that it could not and does not.  Fangyi
said if it's physical behavior of the device, then it's our problem.  If not,
then it's not our problem.

Walter said COM was developed by Richard Mellitz as a "cheap", no simulator way
to get a SNR figure of merit at the decision point.  Walter and Fangyi said COM
uses generic equalization and prescribed optimization algorithms.  AMI can
model specific devices, noise, jitter, etc., that COM doesn't handle.  Wei noted
that eye width and eye height metrics were added to COM 2.95.  Wei, Walter, and
Fangyi said the tool and user could decide if they wanted to disable the Tx
equalization for certain devices.

The group consensus was that no changes to AMI are needed at this time.

BIRD213.1 draft 12, PAMn:
Walter shared draft 12.  He recalled that in the previous meeting we had decided
to remove this sentence on page 288, which Arpad had pointed out was ambiguous
and confusing:
  ... it is assumed that the electrical interface to either the driver or the
  receiver is differential and will have ...

To resolve Arpad's issue, we had first decided to change it to the following:
  ... it is assumed that the electrical interface to both the driver and
  receiver is differential and will have ...
  
Then we had removed the entire sentence.  Walter said that on further review he
thought we needed to restore the sentence to introduce the possibility of more
than 2 levels, which was discussed in the following paragraph.  Walter also
copied the preceding paragraph into both the "replace" and "with" sections to
provide more context to the proposed change.

Fangyi proposed some cleanup to the language in the Other Notes for PAM_Offsets
describing the sampling time of the kth eye.  Walter made the changes.

Walter recalled that Ambrish had suggested that instead of Tables we use String
Type parameters whose values are space delimited lists of numbers for
PAM_Offsets and PAM_Thresholds.  Ambrish agreed that this was his suggestion.
Walter said he had no strong objection to this.  He said they had used the same
technique in Model Specific parameters in the past, for example to return
poles and residues in peaking filters.  However, he thought the preferred IBIS
solution would be to use a Table.

Walter said he would send out a draft 13.

- Ambrish: Motion to adjourn.
- Curtis: Second.
- Arpad: Thank you all for joining.

AR: Walter to send out BIRD213.1 draft 13 incorporating today's changes.
-------------
Next meeting: 22 February 2022 12:00pm PT
-------------

IBIS Interconnect SPICE Wish List:

1) Simulator directives
