IBIS Macromodel Task Group

Meeting date: 12 October 2021

Members (asterisk for those attending):
Achronix Semiconductor:       Hansel Dsilva
Amazon:                       John Yan
ANSYS:                      * Curtis Clark
                            * Wei-hsing Huang
Cadence Design Systems:       Ambrish Varma
                              Ken Willis
                              Jared James
Google:                       Zhiping Yang
Intel:                        Michael Mirmak
                              Kinger Cai
                              Alaeddin Aydiner
Keysight Technologies:      * Fangyi Rao
                              Radek Biernacki
                              Ming Yan
                              Todd Bermensolo
                              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:

- Arpad noted that he had removed the GDDR6x and PSIJ items from the agenda
  because he hadn't received any new updates.  He said he didn't think any were
  imminent and didn't want to leave the impression that anyone owed us an
  update.  Walter said we might want to vote to table the two topics.  Walter
  also noted that the "RDL in IBIS" item in Tabled Topics could be removed.

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

- None.

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

- None.

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

Arpad asked for any comments or corrections to the minutes of the October 5th
meeting.  Curtis moved to approve the minutes.  Arpad seconded the motion.
There were no objections.

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

GDDR6X:
Arpad noted that at last weeks' meeting he had re-introduced a presentation
first given by Vladimir in 2009 on modeling analog Tx effects.  He said he had
wanted to remind us of Vladimir's suggestions, which might help for single-ended
non-LTI simulations.

Walter said he thought he had a solution (which he had reviewed in ATM at the
September 28th meeting).  He said he thought there was no need for a change to
the AMI standard, and it can all be done inside the Tx GetWave with hard coded
information or Model Specific parameters to configure it.  He said he had
implemented an algorithm for taking the input stimulus waveform with 4 evenly
spaced levels and adjusting the amplitudes (based on the symbol level) and the
transition timing (based on which of the 12 possible transitions was occurring).
He said at this point we really need some good SPICE simulation or measurement
data so we can correlate the results from his approach.

Walter said his implementation was not trivial but was something anyone could
do.  He said the only hard part was that if any one of the 12 skew values is
negative then you have to buffer the waveform for up to 1 UI and introduce a
half-UI latency.  He said he thought we had a solution, and we could table the
topic until we have some data to verify the solution or determine that it's
insufficient.

Arpad said his concern with Walter's suggestion was losing the interaction
between the analog model and a channel with ISI, reflections, etc.  He said he
had brought up Vladimir's old presentation because its suggestion was relatively
inexpensive to the simulation engine.  He agreed that we need some more data so
that we can see if Walter's suggestion works.  Arpad also noted that he had
spoken to Vladimir, and his presentation could easily be extended to PAMn.
Walter agreed that we need GDDR6X Tx vendors to implement his scheme or
Vladimir's and see if they work.

Randy said that he had provided some data to Arpad and Walter on the GDDR6X
driver he had.  He said his next step was to look at the equalization that is
available and compounds the issues.  Arpad asked what types of equalization.
Randy said he was still focused on the Tx design and pre/de-emphasis.  Arpad
asked about pre-emphasis as applied to PAM4.  Walter said it's still basically
a two tap FFE, so you just convolve the 4-level PAM4 stimulus waveform with the
two taps, similar to what is done to the 2-level NRZ stimulus waveform.  The
technique is frequency dependent, not really amplitude dependent, so you can
apply the same approach to the PAM4 stimulus waveform.  Fangyi agreed that to
first-order the pre-emphasis works the same way for NRZ and PAM4.  He said you
may then run into non-linear effects, for example something might be affected
by the symbol level, so we consider that as part of the non-linear effects.

Walter noted that Fangyi has a presentation planned for the upcoming Japan IBIS
Summit on November 12th.  Fangyi agreed that his presentation includes modeling
Tx non-linearities.  Fangyi briefly reviewed several slides from the upcoming
presentation, and the group largely focused on slide 9 - RLM (level separation
mismatch).  This slide also described a scenario in which the ideal 4-level AMI
PAM4 stimulus waveform can be modified by the Tx GetWave function.  The model
itself can remap the 4 signal levels to different voltages.  This could model
some of the non-linearities from the Tx.

Arpad said his concern with this approach was that you don't know how the
modulation is working in real time.  With a regular time-domain SPICE simulation
you can capture the interaction of the buffer trying to drive a particular
symbol level and dealing with ISI, reflections, etc., in the channel.  Fangyi
said this is an AMI model, so any non-linearities that are handled exist purely
inside the device isolated from the analog channel.  Arpad said that at DC the
four levels are determined by voltage dividers between the buffer I/V curves,
the channel, ODT, etc., but in real-time even if the buffer is driving the same
symbol level (not transitioning) it has to deal with reflections, etc.  Walter
said that AMI fundamentally assumes the Tx impedance is the same at all four
driving levels.  When we do a step (impulse) response to characterize the analog
portion, we are fundamentally assuming the impedance is constant in time.  If
it is, then all the information about reflections, etc., is captured in the step
response.  Walter said Fangyi's presentation looked like good work, and he
thought they had similar methods for dealing with the non-uniform level
distribution from a PAM4 Tx.

Randy asked how the proposed remapping of the voltage levels worked.  The
actual levels would be very system dependent (I/V of the buffer, resistance
of the channel, ODT, etc.).  Fangyi said the levels are measured at the output
of the Tx.  He said what is corrected for (slide 9) could be considered as an
"average" effect.  You have a constant level mapping independent of the channel
load.  He suggested that one way to deal with a channel load based dependency in
the level adjustments would be to have a set of mappings depending on the load
conditions.  Randy replied that the model might then have a Model Specific
parameter so the user could specify the ODT, for example, and help the model
decide how to adjust the levels.  Fangyi said AMI fundamentally assumes the
analog system is linear.  If the mapping depends on channel load, then AMI may
not be able to fully handle it right now.

Bob asked what we meant by non-linear Tx, what part is non-linear?  Walter said
he considers it a non-linear Tx if the impedances are different at the 4 driving
levels.  He said the impedance is not necessarily what causes the non-uniform
distribution of the 4 Tx output levels.  He said there might be additional
electronics/logic in the driver that decides what voltage level is to be driven.
The selection of that voltage level might be non-linear, but he would not
consider that a non-linear driver if the 4 Tx impedances were the same.

Bob asked if the rising and falling edges were different.  Walter said yes,
there are 12 possible transitions, and each can have different edge rates and
different skews.  AMI assumptions say they all have the same rise/fall edges and
there are no skews between them.  In the data Randy and Arpad had shared in
their DesignCon paper, they had showed edges into ideal loads and you can see
5 to 10ps differences in rise times and skews, but those were into ideal loads.
Walter said he thought that in reality you'd be driving into a channel that is
10 dB down at Nyquist and 40 to 60 dB down at the edge rate.  At the Rx end of
the channel, those differences in Tx transitions may be filtered out and not
very important.

Fangyi agreed with Walter that we probably need more simulation data to assess
the techniques and answer Randy's question.  Randy said GDDR6X data would be
harder to share, but upcoming standards like Graphics 7 would be much more open
standards.

Walter asked if GDDR6X were primarily for video, and if you could accept a
higher BER in video?  Justin said there is a significant difference in required
BER between DDR5 at 1e-16  and GDDR6, where recommendations might be more like
1e-11.

- Walter: Motion to adjourn.
- Bob: Second.
- Arpad: Thank you all for joining.


-------------
Next meeting: 19 October 2021 12:00pm PT
-------------

IBIS Interconnect SPICE Wish List:

1) Simulator directives
