For some reason, computer people presented with mathematical tasks
always like bigger hammers. An amazing number of programs still
do numerical integration by rectangular summation instead of Simpson's
Rule -- and then use ten or more times as much granularity in an
attempt to make up the difference.
OF COURSE if you're relying on a piecewise-linear approximation you're
going to need a gazillion points. The solution isn't to use more points,
it's to get smarter and use a better interpolation. The graphics gang
figured this out a while ago, and THEY do it in hardware at video data
rates.
> ------------------------------------------------------------------------
> ------------------------------------------------------------------------
> -------------
> Marc Humphreys
> Nexabit Networks, Inc.
> 508-460-3355 x236
> e-mail: mhumphreys@nexabit.com
> ------------------------------------------------------------------------
> ------------------------------------------------------------------------
> -------------
>
> > -----Original Message-----
> > From: Al Davis [SMTP:aldavis@ieee.org]
> > Sent: Wednesday, June 09, 1999 3:10 AM
> > To: ibis@eda.org
> > Subject: Re: [SI-LIST] : IBIS Subcommittee Formed to Review Spice
> > to IBIS
> >
> > jonp@pacbell.net wrote:
> > > I have actually been faced with the problem that even with software
> > to remove
> > > extra points I cannot
> > > create IBIS models to the accuracy that I desire with the 100 point
> > limitation.
> >
> > I don't believe even 10000 points would provide the accuracy you
> > desire,
> > because the limits to accuracy are usually elsewhere.
> >
> > For example ....
> > 100 points chosen to be uniformly spaced in voltage, modeling a 5 volt
> > rising waveform, provides a data point every .2 volts. Assuming that
> > the interpolation is completely bogus, this leads to a worst case
> > error
> > of .1 volts. In reality, interpolation is much better than that, and
> > even this choice of data points is far from optimal, so the accuracy
> > is
> > much better than this. Remember, this isn't DC, but the dynamically
> > changing voltage during a high speed transition.
> >
> > Some questions ....
> >
> > Is your Spice model that accurate? (I don't think so.)
> >
> > If you measure one real unit with perfect accuracy, and compare to the
> > next one, are they this close to each other. (I don't think so.)
> >
> > Are your instruments this accurate? (Don't forget that you change the
> > reading when you attach the probe.)
> >
> > What if the real load is different from the test load? Since the
> > driving impedance is nonlinear, and not specified anywhere, even with
> > perfectly accurate and complete data for the nominal load, all bets
> > are
> > off for anything else.
> >
> > Sure, the voltage is right, into a resistor, but what about
> > reflections
> > and noise?
> >
> > How accurate is the rest of the simulation?
> >
> >
> > There are other ways to improve the accuracy. One obvious one is to
> > carefully select the data points (as Kellee said). Another is to add
> > another VT table with a different load. (As Kellee also said) The
> > third table does much more to improve accuracy than more points will
> > do. Tying the load resistor to a middle voltage does a good job at
> > modeling the overlap.
> >
> >
> > I wouldn't object to removing the limit entirely, because whatever the
> > limit is, most files will use it, without any valid reason. With no
> > limit at all, it is up to the modeler to decide what trade-offs to
> > use.
> > Maybe with nothing in the spec saying how many points to use, we might
> > actually get some good 20 point models, that are more accurate than
> > the
> > 256 (or whatever) point models that we will get if the limit is
> > raised,
> > and the capability is there for those rare cases when you really do
> > need
> > the extra points. So, maybe the answer is to remove the hard limit,
> > but
> > add a comment that tables with more than 50 points are rarely worth
> > the
> > extra space.
-- D. C. Sessions dc.sessions@vlsi.com