Re: Models & EDA Vendors

Charles,

Generally, all models from component vendors, whether IBIS,
Quad, or Spice, should be considered suspect until used and
correlated against actual in-circuit operation. Oftentimes
models are not given the care that they deserve at the IC vendor.
They are often "extracted" by someone who doesn't understand
the problem. And the problem is multifold ....

Are the IV curves and waveform tables modeled accurately? If
care is taken, the accuracy of an IBIS model can correlate with
high precision to original HSPICE simulations or physical measurements.
If not, you might be in the right ballpark, you might not.

ICX/Zeelan models are based upon physical measurements, and so have
highly accurate characteristics. However, these are not worst case
measurements. They are based upon a sampling of typical
silicon. Care must be taken by the design engineer to insure that
worst case behavior is taken into account. However, in the absence
of manufacturer's models, these are the best you'll get ... and they
are very good. I use ICX/Zeelan models in my work when
manufacturer models are not available.

Are packages modeled correctly? In order to correctly model package
effects with high performance silicon devices with sub 500 ps
rise times, it is necessary to model the package in detail. This
usually requires a 3D field solver and results in numerous sections
of transmission line. Detailed package analysis by vendors is
rarely provided. Intel is an exception to this rule. There may be
other vendors which also provide this data in Ibis or Quad format, however,
I am aware of few. It is generally next to impossible to get this data
from a manufacturer in any format, even Spice.

So are we totally out of the ballpark for most manufacturers? Well,
that depends. If you are concerned about tight timing margins
in the sub 500 ps range, then highly accurate model and package
characterization are absolutely necessary. You must work with
your chip suppliers to provide you accurate data. This often takes
multiple iterations to "help" them understand your problem.
Intel is an example of a vendor who has "generally" taken care in
characterizing their devices for worst case margin analysis.

However, if you are concerned about signal integrity effects alone,
(non-monotonicity, overshoot, undershoot, ringback, and crosstalk)
and their effect upon correct device operation and timing, and have
a bit of timing margin to spare, then the IBIS models which are
available from ICX/Zeelan, other independent model vendors, and
IC manufacturers will fit the bill. Using these models one can
perform very good board signal integrity analysis.

In my work, I use both options. I am often called upon to analyze
systems that have sub 500 ps and even sub 150 ps timing margins.
In these cases, I work closely with the manufacturers to obtain
IBIS and HSPICE device and package models. I often have to
perform my own HSPICE to IBIS conversions and correlations in order
to "know" that the data I am getting from simulations are accurate.

In other cases, I perform analysis on designs that are not pushing
the state of the art in performance, but are pushing the density
and complexity boundaries of a pc board. Here I am concerned about
noise, overshoot, undershoot, crosstalk and non-monotonicity
of asynchronous signals and clocks. I do complete board analysis
across worst case device corners. However, I usually have at least
1 ns or more of timing margins on most busses. In these cases
I have had incredible success with using manufacturer's, and
ICX/Zeelan based IBIS models. I still have to check them for
reasonableness, and I have to run them through IBIS parser
checks to ready them for my ICX simulation environments, but
these are minor nuisances compared to having no simulation
models at all.

I hope my comments have helped.

Regards,

Scott

"Charles W. Martin" wrote:

> I've recently seen a demo for MGC's ICX tools (Including Tau for timing
> verification). There software uses native IBIS models for timing and
> signal integrity estimations on the fly, and simulation.
>
> ICX mentions a library of 10K parts, as a first tier, afterwards there
> are two-tiers which cost the end-user progressively more for ICX to develop
> ibis models.
>
> ICX mentions that they are working with ic vendors providing them with
> tools which would allow them to develop accurate ibis models themselves.
> I'd like some feedback from anyone (ICX users, IC Vendors, etc..) who
> have worked with ICX on developing IBIS models.
>
> Secondly, we've found that many of the parts we'd like to simulate with
> don't exist in their library. While it's expected that newer components/
> technologies might not have models, some of the models that are lacking
> are typical everyday parts/technologies. What's the preferred approach
> to obtaining and verifying accuracy of these parts?
>
> Lastly, IBIS is a great idea, and the people behind the scenes really
> deserve a round of applause. However, I'm interested in people's opinions
> on how widely available _accurate_ models are from component vendors.
>
> Thanks for your feedback,
>
> Chuck
>
> Charles Martin
> Cabletron Systems, Inc.
> EDA Tools
> cmartin@ctron.com
> Phone: (603) 337-2973
> Fax : (603) 337-1764
>

--
___________________________
Scott McMorrow
Principal Engineer
SiQual
mailto:scottmc@teleport.com
___________________________