[SI-LIST] Re: package SSN model accuracy requirements

One thing I have to agree with your, none of us have any time left to get
into this infinite loop of endless "my **** is better than your ****". As
long as I am getting encrypted HSPICE from vendors today, I could care less
Joe engineer thinks IBIS is all he needs. 

>Your assessment that IBIS can't do it at all is plain wrong.

"On IBIS and SSN: many in the industry recognize that, for effects such
as SSN, traditional IBIS (3.2, 4.0) today does not include enough data
to provide a consistent power integrity result between tools."

I didn't say the above, Michael did.

And for the last time, NO ONE in Si-list response saying he/she is doing SSO
analysis with IBIS today. How many tools out there really support the latest
spec of IBIS 4.x anyways ?

You can't dance around the encryption issue because that was one of the
biggest reason IBIS exist in the first place. 

Call it SPICE or AMS, at the end of the day, you can't be accurate enough on
your model without someone in your chain of command thinks that's too much
IP and you need to encrypt the information.

And do you really think all favors of Verilog-AMS, VHDL-AMS with different
encryptions for different vendors will really fly ? With your busy schedule
between real design work and your IBIS meetings ?

Call me crazy, I've design plenty of big irons with SPICE and I never felt
its slowest hinder my work, I am more worry about garbage models that lead
me to incorrect results. If tomorrow you can show me an AMS model that can
simulate SSO on drivers accurately, I'll take it. But so far Gary's example
is no more than a two pass simulation with digital verilog simulation
feeding into a SPICE level=1'ish transistor model and S-parameter. And your
example is just a mix of couple of IBIS drivers pumping current out at
different time. None of this address the problem with SSO noise. Unless we
think that's not any issue, you haven't even start to address the problem. I
can take a few standard IBIS driver models, run a digital verilog
simulations and pump the different drivers at the output based on the
verilog timing and I can get the same results. What's so special about it ?
It doesn't address the problem of SSO noise. I think at this time you and
Gary will probably say "but hey, I can modify this and that and it will do
the job". Here lies the problem, if you and Gary are already the bleeding
edge of AMS users and both of you still can't generate the SSO model, how
long do you think the rest of the app engineer crowd can catch up the
methodology ? I hope its not seven years.

I am really glad Al actually make his presence in Si-list. May be he can
share with us how long since he did those SSO analysis on IBM TCM modules.
To see us little disciples still struggling to figure it out will probably
make him laugh. But it also tells us how slow the entire industry is moving.
A promising technology may be great but let's not loose sight we need to
address a lot of problems today.

Arpad, you should know better I can make any methodology you throw at me
works. SPICE BTDT, behavioral BTDT. AMS, no problem, I am sure I can make it
work also. But can you say the same about the Joe app engineer of the
majority of the industry ? Why can't there be a SINGLE IBIS SSO model
generated and people are using it today ? Is there no SSO problem today ? 

Up to this point IBIS has some reasonable strict definition, I-V curves,
package parasitics etc. It may not be good enough for SSO but at least we
all know where/how it comes from. If tomorrow everything becomes this
magical AMS model and all users see is a bunch of equations, who is going to
guarantee the accuracy or the methodology used to abstract it from SPICE
(which I believe is what 100% of Si I/O designers use) to AMS ? I've ask you
previously already, how do you determine what kind of abstraction is good
enough for what kind of application ? Can it even be defined ? If the answer
is no, who is going to guarantee the level of accuracy ? HSPICE may not be
good, but at least I've never heard of any Si vendor unwilling to correlate
their process with it because, like it or not, it is considered a de facto
industrial standard. Can you say the same about all those AMS abstractions
that will exist in the future ? 

I am a practical man, if tomorrow AMS is the heat it claims to be, I will be
the first to jump on it. But you have to show me a single case where the
average Joe app engineer can hand out an AMS model with SSO model first. Can
you ? I am not talking about "I think it can be done this way", I talking
about "here is the I/O with SSO model, here is a scope picture of actual
measurement of SSO of the I/O and here is an AMS simulation that matches
it".

I will be waiting for that picture. 

And how long do you think the rest of the industry will show me that kind of
pictures ? Please don't tell me another seven years.

-----Original Message-----
From: Muranyi, Arpad [mailto:arpad.muranyi@xxxxxxxxx]
Sent: Thursday, March 24, 2005 2:20 PM
To: Chris.Cheng@xxxxxxxxxxxx; si-list@xxxxxxxxxxxxx
Subject: RE: [SI-LIST] Re: package SSN model accuracy requirements


Chris,

Here are my responses.

When we started IBIS, our main goals were not SSO simulations, at least
not the way it is done today.  That's why we put only a partial and
crude solution for it into the IBIS specification in the form of the
[Pin Mapping] keyword.  It did provide some capability for SSO, exactly
with the same thinking you are saying: "at least give you a chance to
predict the problem".  Your assessment that IBIS can't do it at all is
plain wrong.

The thinking in those days was that if it becomes necessary to do
something better, we will add it to the spec.  This is actually
happening right now in the form of BIRD95.

There may be many reasons it took this long to get this far.  One
may be that instead of doing something about it, people just kept
complaining about it.  Another could be that none of us in the IBIS
Open forum are "professional standards attendees", we all do this
above and beyond our normal everyday work on a volunteer bases,
which leaves little time for making improvements on the specification.

The same goes for the receiver modeling problem.  When IBIS was started,
there was no need for doing that, so we didn't put it in the spec.  In
those days most everything was measured at the pin of the device.  Later
we started to move the measurement point to the inside of the package,
the die pad.  This was still doable with IBIS.  Timing measurements at
the output of the receiver are not widely done in the SI world yet.
Since IBIS is mostly demand driven, this hasn't been addressed yet
in the spec.  However, with the *-AMS capabilities we are not going to
need to change the spec any more if it becomes necessary to model the
receiver.  As I stated it before, in *-AMS you can model things any
whichever way you want to.  You can make an exact SPICE equivalent if
you like, or you can write a behavioral block if that is your choice.

Regarding your original question, you have already gotten two responses,
so I will not go into that any further.

Regarding the rest of your questions:

a)  I don't want to go into encryption, because that is a separate
subject all together.  However, regarding my company's secret in
English, or Russian, or Martian, I disagree some.  Think about a 
simple voltage divider example.  You can write a Thevenin equivalent
for it using one voltage source, and one resistor.  The resistor's
value will be the parallel combination of the original circuit's
resistors, and the voltage source will have the value of the voltage
division the divider gives you.  If you don't know anything about
the original circuit and I gave you only the Thevenin equivalent,
are you going to be able to tell me what the two resistor values are
in the original circuit and what voltage they are connected to?
That's how behavioral equivalents can be useful to hide my
company's secret.  I agree, it may not be always possible to hide
everything like that, and the lawyers may find that my Thevenin
equivalent is IP, but that is a different story.

b)  Asking how AMS can handle gate modulation is equivalent to ask
how the C++ language can handle your favorite software projects.
By the way, the HSPICE B-element has done this for years now, try
the "spu_scal" and "spd_scal" parameters of the B-element.  It just
hasn't been made official by the IBIS specification yet, because
there was not enough demand for it up to now.  However, you can
look at BIRD97 and the comments Katja wrote recently how this 
would work and how it could potentially become part of the spec.

Regarding your concerns about speed when the simulation deck is
dominated by interconnects, you can actually combine all your
interconnect pieces into one big S-parameter block (i.e. make
a behavioral model for them), and then your simulation is going
to go orders of magnitude faster.  Same principle as combining
the thousands of transistors into a behavioral buffer block...

The equalization stuff or receiver modeling doesn't have to be done
with multi dimensional lookup tables to be behavioral.  This is
exactly where the AMS languages provide a tremendous capability.
You could write FIR filters, decision feedback loops, etc. with its
digital equation capabilities, which would execute lightening
fast, because they are event driven, and don't have to be
evaluated at every single analog time step (iteration).

c)  Again, I do not want to go into the subject of encryption,
it is a different story.  However, when we write *-AMS or AMS,
we are talking about VHDL-AMS and Verilog-AMS.  Quote from the
IBIS4.1 specification:

| "VHDL-AMS" refers to "IEEE Standard VHDL Analog and Mixed-Signal
| Extensions", approved March 18, 1999 by the IEEE-SA Standards Board and
| designated IEEE Std. 1076.1-1999.
|
| "Verilog-AMS" refers to the Analog and Mixed-Signal Extensions to
| Verilog-HDL as documented in the Verilog-AMS Language Reference, Version
| 2.0.  This document is maintained by Accellera (formerly Open Verilog
| International), an independent organization.  Verilog-AMS is a superset
| that includes Verilog-A and the Verilog Hardware Description Language
| IEEE 1364-2001.

Sincerely,

Arpad
------------------------------------------------------------------------

-----Original Message-----
From: si-list-bounce@xxxxxxxxxxxxx [mailto:si-list-bounce@xxxxxxxxxxxxx] On
Behalf Of Chris Cheng
Sent: Wednesday, March 23, 2005 5:27 PM
To: si-list@xxxxxxxxxxxxx
Subject: [SI-LIST] Re: package SSN model accuracy requirements

Arpad,

I can give you an even simpler answer, ever since the existence of IBIS, it
can't model a simple problem like SSO. If you are a customer, would you
choose a method that may be error if it is done wrong but at least give you
a chance to predict the problem if it is done right, or, something just
can't do the analysis AT ALL ? There are many road accidents with cars, but
that doesn't mean you have to walk to work.

Let me repeat the question I've asked many many times, how many of the
member of the Si-list who are NOT just professional standards attendee or
EDA tool vendors and have real responsibility to design system are actually
using a standard IBIS model to analyze SSO today ? I rest my case.

Garbage in garbage out, no matter it is SPICE or IBIS. At the end, whoever
give the model out has the ultimate responsibility. But ever since the
existence of IBIS, it can't address the problem high performance design (SSO
and receiver performance to name a few). You and I know that behavioral
models can address those problems long time ago but look at how long since
the first known successful attempt (if you disagree with me on that
reference, ping me off-line) until the standard committee even catch up with
the idea. Seven years is a very long time, people can earn sabbatical out of
that. Am I supposed to wait for another sabbatical if I want to model my
equalizing receivers ? I've got problems on things I need to ship tomorrow. 

Let's take a closer look at AMS
Here are the claims :

a) It abstract your I/O to protect your IP
Well, Gary's reference seems to suggest every I/O design is as simple as a
university paper so may be your company have no problem handing out the I/O
state machine design. Is that true ?
On the same reference we are led to believe a SPICE level=1'ish model is
really a behavioral model, ok I dig it, but do you think your company
lawyers and design managers will buy that and freely handing it out without
encryption ?

Like I said before, it's like asking "Can you tell me the secret of your
company ?"
Ans : "I can't tell you in English, but I can tell you in Martians (just not
to offend my earthly friends :-D)" 
Is that really possible (besides the fact that there is no Martian)?

b) It is accurate and fast
I still haven't seen any mention of SSO or how AMS can handle power impact
on the I/O, may be SSO is no longer an important issue for GB serial ports ?
How does one handle such modulation by predriver and substrate feedback from
multiple power sources (core power for predriver, I/O power for main driver)
without resorting to those level=1'ish transistor model ? A marketing paper
like what Gary present can carefully craft the example to the advantage of
the simulator but in reality how many interconnect is one simple S-parameter
deck ? I would bet the real customer topology will more like a mix bag of
circuit elements different vendor provide such as chip package, transmission
line, terminators, discrets, connector models. How the speed of AMS will do
when it is overloaded with a lot more circuit elements like R,L,C,
transmission lines and S-parameters by different components ?
Let's focus more on this little problem like a receiver, if you have to use
the output of the receiver to predict a equalization how would you use your
behavioral models to predict that ? Just take a look at those typical crazy
FSB ringback, over-drive specs. Try a few case of real over drive and ring
back cases which depends on common, differential mode, operating point, over
drive, hysteretic etc and try construct a multi-dimensional equation/table
to describe it and let's see how fast you will get. Ask the friends we both
know and have already make pitches in this thread what do they think ? I am
a fair person, show me a real life case and data and I will be convinced.

c) It is standard and everyone support it
I still couldn't figure out whether AMS is VHDL-AMS or Verilog-AMS, which
one we are talk about here ? Can you tell me ?
Does everyone use the same encryption to protect your IP ? If every tool has
a different encryption, what kind of ZBB do you think you will get for ten
different kinds of encryptors from ten different vendors ?
------------------------------------------------------------------
To unsubscribe from si-list:
si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field

or to administer your membership from a web page, go to:
http://www.freelists.org/webpage/si-list

For help:
si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field

List FAQ wiki page is located at:
                http://si-list.org/wiki/wiki.pl?Si-List_FAQ

List technical documents are available at:
                http://www.si-list.org

List archives are viewable at:     
                http://www.freelists.org/archives/si-list
or at our remote archives:
                http://groups.yahoo.com/group/si-list/messages
Old (prior to June 6, 2001) list archives are viewable at:
                http://www.qsl.net/wb6tpu
  

Other related posts: