Subject: Re: removing non-monotonic points?!
From: Scott McMorrow (scott@vasthorizons.com)
Date: Thu Feb 07 2002 - 18:23:34 PST
Andrew,
I was not talking about a hypothetical when I suggested that removal of
and/or smoothing non-monotonic points in the output I/V curves is a
practical method of modifying models. I am talking about the actual
reality in the simulators which we use. We have removed points from
IBIS models and correlated the results to the original HSPICE
simulations with the following three IBIS simulatiors
HSPICE b-element
Mentor Interconnectix
Cadence SpectraQuest.
In all cases, the removal and smoothing of points which are
non-monotonic ( generally below the ground rail, but sometimes above the
power rail) have absolutely no effect whatsoever on the simulation
results in these simulators. When we do this, for models which we
produce, we do perform correlations with the IBIS accuracy metric. We
have data on nearly 500 buffers which we have created models for which
illustrate this point. If our correlation to HSPICE is not at least 97%
into a reflected wave test environment at both the driver output and at
the end of the line, we start over and refine the extraction process.
In nearly 80% of all buffers which we extract IBIS Models for, the
correlation is generally more on the order of 99% or better, with only
very slight visual differences in simulation waveforms.
There may be simulators where this smoothing of the data is a problem,
but I've not found one so far. There may also be some types of outputs
where correlation mismatches could occur, however we have not seen one
yet in all our years of experience in creating IBIS models from the
original HSPICE.
Generally, the excessive undershoot which you see at the pins of a
device are not occuring at the die. Often what is measured at the pin
has a huge offset due to SSO and SSI (Simulaneous Switching Inputs),
which IBIS modeling does not account for accurately, anyway. If the
clamp is weak, then we often find that there are no non-monotonic
regions (until device breakdown) in which case the full effect of the
undershoot would be modeled correctly, in your case of -2.5 V negative
overshoot. I'd argue strongly that for strong negative overshoot, the
modeling of clamp diodes in IBIS and in most HSPICE models is not up to
the task anyway.
regards,
scott
-- Scott McMorrow Principal Engineer SiQual, Signal Quality Engineering 18735 SW Boones Ferry Road Tualatin, OR 97062-3090 (503) 885-1231 http://www.siqual.comIngraham, Andrew wrote:
>As I've spoken my mind about this recently, I figured I should throw my >two cents in. I disagree that I-V points more than 800 mV beyond the >rails are meaningless. > >Even though a simulator may not CONVERGE on a point that is 2V beyond >the rails (because of stiff clamping), that doesn't mean that the >simulator didn't TRY that point. If the I-V data at that point in the >model is totally out to lunch, say zero, then you wouldn't expect the >simulator to figure out it needs to get back into a more normal >operating region. > >That's one reason why I-V tables should never be truncated, even if you >can't measure the current that far out without frying the part. The >data out there needs to be reasonable enough ... by extrapolation if >need be ... to guide the simulator back toward reality. Otherwise it >might actually converge way out there, or may give up and abort the >simulation. > >Also, there are plenty of duff SPICE models that have unusually stiff >clamps because they forgot to include any diode resistance. With these >models you might never find a simulated result that is more than 800 mV >beyond the rails, no matter how hard you try; whereas the real devices >are not that ideal. I have seen negative overshoot on a few devices go >as far as about -2.5V ... although this is at the package pin, not on >the die. Suffice it to say that some parts clamp less well than most. > >Regards, >Andy > > >
This archive was generated by hypermail 2b28 : Thu Feb 07 2002 - 18:45:18 PST