Kumar,
Not true. Optimizing the Tx to open the eye at the input to the Rx is just
wrong at date rates above 6 Gbps, and I have the data to prove it.
It probably is OK when we first put AMI together and the state of the art
was 3 Gbps and there was just a two tap Tx and a one tap DFE and very
limited CTLE.
Walter
From: "" <dmarc-noreply@xxxxxxxxxxxxx> (Redacted sender "kumarchi@xxxxxxxxx"
for DMARC)
Sent: Thursday, June 11, 2015 6:12 PM
To: wkatz@xxxxxxxxxx; Arpad_Muranyi@xxxxxxxxxx; ibis@xxxxxxxxxxxxx
Subject: Re: [ibis] Re: Flexibility of a specification
Walter:
Tx optimization is very useful and in the absense of backchannel support
works very nicely with proper constraints on post cursor.
also rx devices are not so brittle that they need eda tool to figure out
their adaptation settings. The devices are very sophisticate, robust and
will not under wide variety of starting points
_____
From: Walter Katz <wkatz@xxxxxxxxxx <mailto:wkatz@xxxxxxxxxx> >
To: Arpad_Muranyi@xxxxxxxxxx <mailto:Arpad_Muranyi@xxxxxxxxxx> ;
ibis@xxxxxxxxxxxxx <mailto:ibis@xxxxxxxxxxxxx>
Sent: Thursday, June 11, 2015 5:50 PM
Subject: [ibis] Re: Flexibility of a specification
Arpad,
You are correct. If we are not talking about communication from the Rx to
the Tx, then the EDA tool is controlling the configuration of the Tx and the
Rx. This is one of the modes that we want the optimization BIRD to support.
Currently EDA tools need to do a blind search (or we prefer an intelligent
blind search) to run the simulations with a range of Tx tap equalizations.
This is why the original concept of the Tx model optimizing itself was
always fundamentally flawed. We cannot do anything with old models but we
could (should) add a reserved parameter to the Tx so that the Tx DLL can be
run without risking the Tx changing its equalization based on the Impulse
Response of the channel that is input to the Tx (as I proposed in the
Redriver Flow clarification BIRD that is currently tabled in IBIS_AMI.
Walter
From: ibis-bounce@xxxxxxxxxxxxx <mailto:ibis-bounce@xxxxxxxxxxxxx>
[mailto:ibis-bounce@xxxxxxxxxxxxx] On Behalf Of Muranyi, Arpad
Sent: Thursday, June 11, 2015 5:28 PM
To: ibis@xxxxxxxxxxxxx <mailto:ibis@xxxxxxxxxxxxx>
Subject: [ibis] Re: Flexibility of a specification
Walter,
Regarding: “…in no way precludes the Rx from setting the Tx equalization.
Thus the CTLE can be optimized with whatever Tx equalization the Rx model
chooses. I would expect that an Rx optimization algorithm will set the Tx
equalization to off (or the minimal amount of Tx equalization that will
allow the Rx to evaluate an eye), then optimize its CTLE and DFE, and then
make incremental changes to the Tx to fine tune the channel. So for those of
us evaluating the various equalization messaging methods one would want one
that is capable of setting the Tx equalization to either off, low, or one of
a number of presets, and then be able to change the Tx equalization
incrementally.”
this sounds like you have the Backchannel BIRD (147) or its derivatives
in mind which is still in the future (if it will ever happen). I was
talking about the existing spec and existing models, and this was a
simple illustration with one example from the spec for why I think the
spec is inflexible.
Aside from that, you missed the point of the thoughts I had in mind
when I started this thread. The point I am trying to make is that we
will continually run into limitations if we use the spec to describe
what and how people should do. Consider plain old SPICE. Does the
SPICE language tell you that you can only write subcircuits with a
certain predetermined “footprint”? Does the SPICE language tell you
that you can only simulate the subcircuits you are allowed to write in
a certain order? Obviously not. That’s what I am suggesting we should
also do in IBIS-AMI (or IBIS in general). Give people a language, let
them write any function they want, and give them a simulator that allows
them to execute these functions in any order.
Disclaimer: There are my personal views, not my employer’s views. I
may say different things when I represent my employer’s views.
Thanks,
Arpad
========================================================================
From: ibis-bounce@xxxxxxxxxxxxx <mailto:ibis-bounce@xxxxxxxxxxxxx>
[mailto:ibis-bounce@xxxxxxxxxxxxx] On Behalf Of Walter Katz
Sent: Thursday, June 11, 2015 4:01 PM
To: ibis@xxxxxxxxxxxxx <mailto:ibis@xxxxxxxxxxxxx>
Subject: [ibis] FW: Re: Flexibility of a specification
All,
I have seen 18 e-mails yesterday on IBIS-AMI. They indicate the following
things to me:
1. There are many of us (users and EDA vendors) frustrated with the
quality of AMI models delivered by IC Vendors
a. Models are not interoperable
b. Models do not work on all EDA platforms
c. Models do not work on all computer hardware
d. Models are delivered DOA
e. Models do not comply with the standard
2. There are many users frustrated with the fact that EDA platforms do
not give the same result
3. The IBIS spec does not have a cookbook that tells model makers how
to write models
4. Some of the e-mails are written by people who do not understand
what AMI modeling is and what it is not
5. …
What is IBIS-AMI
There is the IBIS 6.0 specification and a paper we recently gave at
DesignCon
http://www.designcon.com/santaclara/scheduler/session/understanding-ibis-ami-simulations.
But let me describe here how simple it is. AMI models are executable code
that links in directly to the EDA simulation tool as a Dynamically7 Loaded
Library (DLL on Windows or a Shared Object on Linux). There are three
functions in the DLL: AMI_Init, AMI_GetWave and AMI_Close.
* The AMI_Init function initializes the function and can output
results that an EDA tool can do statistical analysis.
o The input to Tx AMI_Init is an Impulse Response of the channel and
configuration of the model as described in an ASCII .ami file that is
supplied with the DLL.
o There is no magic to create an Impulse Response of the channel. It can
be done by simulating a Step Response with the IBS model for the Tx and Rx
and the interconnect (channel) between the Tx and Rx. The time derivative of
this Step Response is the Impulse Response. There are other methods of
generating this Impulse Response in the frequency domain as well. They
should give the identical result (within the numerical accuracy of each of
the methods). This Impulse Response (or a Pulse Response that can be
trivially generated from an Impulse Response) is exactly the input to
Stat-Eye, COM, or any SerDes MatLab based tool.
o There is no magic to create the configuration that controls the
operation of the DLL. This is explained in the IBIS 6.0 spec.
o The output of the Tx AMI_Init function is an Impulse Response modified
by the Tx equalization (usually but not necessarily an FFE filter)
o The input to the Rx AMI_Init function is the Impulse Response output of
the Tx AMI_Init function, and the configuration of the Rx model as described
in the Rx .ami file.
o Both the Tx and Rx .ami files describe the controls that the user has in
configuring the silicon itself, and should correspond either directly or
with some mapping software into the registers that the hardware used to
configure the silicon.
o The output of the Rx AMI_Init function is an Impulse Response at the Rx
latch (decision point) that the EDA tool can use to generate a Statistical
Eye, and information that the EDA tool can use to generate a Clock PDF
function.
o The EDA tool is responsible for generating bathtub curves, BER, and
other analysis of the Rx output.
* The AMI_GetWave function is used to do time domain simulations.
The input to the Tx AMI_GetWave function is a stimulus waveform with a
well-defined format, the output of the Tx_GetWave function is a waveform
which includes the Tx equalization. This is convolved with the Impulse
Response of the channel to generate the waveform input to Rx AMI_GetWave.
The output of Rx AMI_GetWave is a waveform at the Rx latch (decision point),
and clock times that can be used to sample the waveform. The EDA tool is
responsible for analyzing the Rx outputs.
* The AMI_Close function is used when the simulation is complete to
free memory.
This is all there is. The inputs and outputs requirement are precisely
defined by the IBIS 6.0 standard. All is not perfect however. A model can
support just statistical, just time domain or both statistical and time
domain processing. This is particularly problematic when a Tx model does
not support statistical.
There are other special cases – like repeater models and retimer models that
are not discussed here.
As for comments about the flexibility of the specification, both SiSoft and
Keysight demonstrated prototype implementations of PAM4 in their AMI
simulators at DesignCon. It only took us a couple of weeks to resolve our
differences and write a BIRD (172) to enable PAM4 AMI modeling. Dotting the
I’s and crossing the T’s took another month or two. The resulting BIRD is
ready to be approved at this week’s Open Forum. IBIS would be well served by
approving this BIRD and including it in IBIS 6.1. IBIS should commit itself
to being responsive to the needs of the IC Vendor, User and EDA Vendor
communities.
What are the Problems
The biggest problem that we see are non-compliant models. It is easy to
write a parser for the .ibs and .ami ASCII files to insure that they comply
to the IBIS specification, it is harder to test for the actions of the
executable to be compliant. Engineers who develop IP and use tools such as
MatLab often do not have the expertise to write quality C code, they do not
necessarily have both 32 bit and 64 bit environments to build on, and they
use tools to develop their models that might have samples per bit or block
size limitations. Writing bug free software is not easy, but just like there
is no excuse for model writers delivering IBIS files that do not pass the
IBIS parser, there is no excuse for AMI model writers from running their
models on Windows and Linux 32 bit and 64 bit computers exercising every
possible value of each of the AMI parameters that are input to the model,
and making sure that the models to not crash, and give expected results. It
is inexcusable for a model writer to deliver a model without documenting the
inputs and outputs and the models limitations.
The next problem that is tarnishing the AMI modeling committee is that EDA
Vendor tools do not use the AMI models correctly. See for example
http://www.designcon.com/santaclara/scheduler/session/ibis-ami-model-simulations-over-six-eda-platforms.
The bottom line is that Users need to reject IC Vendor parts when the IC
Vendor AMI models do not pass a set of usability requirements. Nothing gets
an IC Vendors attention better than the Golden Rule, he who has the Gold
Rules.
Both SiSoft and Cadence have supplied the source code for a program that can
exercise an AMI DLL. It is not big deal to write scripts that can exercise a
DLL with a full set of inputs.
One Should Not Confuse Analogue Simulation with AMI Simulation
Generating the Impulse Response is a different problem then generating the
AMI simulation. For example, the following statement is misleading:
“A number of folks that I communicate with have expressed great frustration
with the use of IBIS-AMI models in their simulators of choice. Are there
examples and a user guide which go beyond the specification?”
One should not expect or require that a single simulator can both analyze
the channel to generate the Impulse Response and also do the AMI simulation.
These are separate problems. There are simulators that can do both, but
there are simulators out there that do one well and the other either not at
all or poorly.
So a SPICE simulator can be IBIS compliant, but not AMI compliant. It is
certainly valid to use that SPICE simulator to generate the Impulse Response
of the channel, but do not expect it to be able to do the AMI simulation.
Optimization Confusion
There were several e-mails relating to optimization that indicate that even
us experts do not agree on what the specification says. E-mails like the
following muddle the landscape and need to be called out and corrected:
I would like to give you an example for the inflexibility of the IBIS-AMI
specification. The words I used before (“canned”) may not have been the
best choice, but let’s not argue over that.
Multiple SERDES experts would say these days that there are times when it
is better to optimize the Rx CTLE to the channel’s IR without including
the Tx EQ effects. Since the CTLE is usually LTI, its algorithm can very
well be placed in the Rx Init function. However, on the bottom of pg. 172
this is what the spec says:
“Under certain circumstances, for example when the Rx AMI_Init function
includes an optimization
algorithm, the impulse response presented to the Rx AMI_Init function must
include the Tx
equalization effects for the optimization to work correctly.”
This statement doesn’t seem to allow the CTLE be optimized without the Tx
EQ’s effects. Do you consider this to be a flexible specification? Do you
see a way around this requirement in the DLL’s algorithms? Do you think
it is OK for EDA vendors to ignore this statement in the specification and
still be considered IBIS spec compliant?
Although Cadence and SiSoft disagree on the details of how messages are
defined that get passed back and forth between the Tx and Rx model, we both
agree that the Rx model controls the Tx equalization. The statement “the
impulse response presented to the Rx AMI_Init function must include the Tx
equalization effects for the optimization to work correctly” in no way
precludes the Rx from setting the Tx equalization. Thus the CTLE can be
optimized with whatever Tx equalization the Rx model chooses. I would expect
that an Rx optimization algorithm will set the Tx equalization to off (or
the minimal amount of Tx equalization that will allow the Rx to evaluate an
eye), then optimize its CTLE and DFE, and then make incremental changes to
the Tx to fine tune the channel. So for those of us evaluating the various
equalization messaging methods one would want one that is capable of setting
the Tx equalization to either off, low, or one of a number of presets, and
then be able to change the Tx equalization incrementally.
32 Bit VS 64 Bit
This is a potentially difficult computer science problem that is simply
solved by the model maker supplying both 32 bit and 64 bit versions of the
model.
Advanced Features – Supporting Innovation
Models that support advanced features do not work on all simulators. This is
not a negative – it is a positive! Take PAM4 as an example. The authors of
the PAM4 BIRD (SiSoft, Keysight, Avago and Xilinx) are developing (and
delivering) PAM4 models to companies that are designing PAM4 systems as we
speak. We initially privately, and then publicly in BIRD 172 defined new AMI
Reserved Parameters that have enabled IC Vendors to move forward developing
models, enabled EDA tool vendors to move forward enhancing their tools, and
users to evaluate channels and make engineering decisions today. Until BIRD
172 is approved, and until IBIS 6.1 (or so) is approved, and until IBIS
supplies a parser that supports these new Reserved Parameters, we are all
supporting these parameters in the Model Specific section of a .ami file.
Converting these IBIS 6.0 models to IBIS 6.1 will simply require that these
Model Specific parameters be move to the Reserved Parameter section of the
IBIS file. Other EDA Vendors are free to implement this PAM4 functionality
as they see fit. So I quote
“Only the 2nd can be blamed on the IBIS committee and only if the complaint
is that they don’t move fast enough.”
Non-Compliant AMI Models
The only solution to deal with model makers who supply models with fixed
samples per bit or fixed block sizes, is for users to not buy those parts
(remember the Golden Rule). If model makers think it is too much of a burden
to write models at any block size and samples per bit, they should submit a
BIRD that makes Samples_Per_Bit and Block_Size reserved parameters, become
dues paying members of IBIS and vote it in the standard. These problems have
been around from the beginning. We even submitted a BIRD to allow a model to
declare Samples_Per_Bit as a reserved parameter that would have required the
EDA tool to do the “Torque Conversion” in the simulator.
Enforcing Better Models – AMI Model Plug Fests!
There were statements like “IBIS should require a Model User Guide”, “IBIS
should supply tools that validate models”, … It is easy to say that every
AMI model should have a Model User Guide, and someone can make such a motion
at this Friday’s Open Forum. Is someone going to write a specification one
what should be in an AMI Model User Guide? Can we write a parser to verify
that a Model User Guide is complete and correct? Who is going to spec a tool
to validate a model? More interesting, who is going to write a spec to
validate an AMI EDA Tool? Imagine organizing an IBIS AMI Plug Fest (like the
PCIeG3 Plug Fests).
IBIS Cannot Do Anything, But You Can
The IBIS organization is most likely forbidden to name specific IC Vendors,
EDA Vendors or User offenders of the standard in order to protect the
innocence of the guilty, and I think this will prohibit IBIS from
effectively addressing these problems. Ultimately it is the user of IBIS
models that need to stand up and do this. I suspect that someone could start
a Blog somewhere and continue this discussion outside of the IBIS umbrella.
Walter
Walter Katz
wkatz@xxxxxxxxxx <mailto:wkatz@xxxxxxxxxx>
Phone 303.449-2308
Mobile 303.335-6156
