Meeting minutes are attached.
Randy
IBIS Macromodel Task Group
Meeting date: 5 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. Randy Wolff took the minutes.
--------------------------------------------------------------------------------
Opens:
- None.
-------------
Review of ARs:
- Walter to send his GDDR6X presentation to the ATM list.
- done
--------------------------
Call for patent disclosure:
- None.
-------------------------
Review of Meeting Minutes:
Arpad asked for any comments or corrections to the minutes of the September 28th
meeting. Jared moved to approve the minutes. Randy seconded the motion.
There were no objections.
-------------
New Discussion:
GDDR6X Signaling with IBIS-AMI
Arpad was thinking about the issue of losing interaction with a model driving
into a channel and dynamic reflections from the channel to the Tx. Arpad
shared a presentation from Mentor Graphics from October 14, 2009 "How to
account for non-LTI of Tx analog buffer in IBIS AMI flow".
Link on ATM site:
https://ibis.org/atm_wip/archive/20091014/vladimirdmitrievzdorov/How_to_account_for_non-LTI_of_Tx_analog_buffer_in_IBIS_AMI_flow.zip
Slide 13: With an additional convolution, one could solve for the non-linear
response of the Tx for GetWave. Vladimir was proposing to take into account
driver non-linearity, and it would not take a lot of extra overhead in
computational expense. Taking into account the Rx in a similar way would be
more difficult, but he was proposing to only take into account the Tx, which is
where most of the non-linear effects occur.
Bob asked about more details on Vt(t). Arpad said Vt(t) is the resulting
waveform calculated in Tx GetWave accounting for all of these effects.
Walter commented that rise time of GDDR6X/7 are very fast, maybe 10% or less of
a UI. Can also look at impedance of pullup and pulldown. The I-V curves of
these buffers, the impedances might look the same at the 4 levels. Most of the
shaping of the waveform at the Rx is due to loss of the channel. You're not
going to see the edge rates, only sine waves at the Rx. If what he is saying
is true, non-linearities are second order effects.
Arpad said he did not dispute these points, but he said that the final voltage
levels of the driver into the channel still seems to depend on these non-linear
effects. That reminded him of this presentation which was trying to solve the
issue.
Walter would like to see more data and models from IC vendors to help determine
the magnitude of the issue.
Arpad said modulating the PAM4 input levels at the input of Tx GetWave, as
Walter suggested last week, would include the time-dependent interactions
between the driver and channel.
Randy commented that the non-linearity of the Tx I-V curves is a trade-off
between device capacitance and linearity. Design decisions are made about how
linear the Tx needs to be. If the I-V curve is available, what would an EDA
tool do with it? Could it solve for DC voltage levels in a system simulation
using those I-V curves, then modulate the PAM4 input levels to the IBIS-AMI
model? Arpad noted this would change the current method of using a single
impulse response capturing the rail to rail swing and assuming evenly spaced
PAM4 modulation levels.
Walter noted the datarate now versus 12 years ago. With a 10GS/s signal and 5
inch channels, you'll get losses, and you want to be sure you won't be getting
a resonance to collapse the eye and design for ISI improvements. Walter noted
the GDDR6X data should exist to see if the levels are mismatched in real
hardware. He'd like to see that data to see how accurate the models really are.
Arpad added that Vladimir's presentation was written at a time when the PAM4
part of the specification was not done. He might ask Vladimir to update the
slides with some notes on PAM4.
- Ambrish: Motion to adjourn.
- Radek: Second.
- Arpad: Thank you all for joining.
-------------
Next meeting: 12 October 2021 12:00pm PT
-------------
IBIS Interconnect SPICE Wish List:
1) Simulator directives