In our lab we are able to measure the s-parameters on just about
any physical package up through 20 GHz, thoug we seldom look above about
5 GHz at present. We then use the captured parameters to perform an
optimization of a candidate topology using either HP MDS or Aplac. Typically
I've seen very good agreement (less than ~5% divergence) between measured
and simulated results up through about 2 GHz. At that frequency the delta
increases. A more complex model could be used to provide less error
between measured and simulated data to a higher frequency, however with
the rise-times currently in use today 2 GHz BW appears to be adequate
for the moment, though applications where a higher bandwidth is necessary
are not far off.
In addition to creating lumped element models based on optimization
techniques, we've also used "black-box" models created by performing a
Pade' approximation on measured s-parameter data. One needs to be careful
when using that technique as we seen cases of instability result from
passive components due to the way the poles and zeros were extracted to
create the ratio of polynomials. I've also found that other users of the
models to be suspicious of "black-boxes" in their simulations and seem to
prefer lumped element models.
Another thing to be wary of when using models based on Pade'
approximations is that certain versions of spice (Hspice up through version
95.1 at least) have problems handling very small and very large exponents
in their polynomial controlled voltage and current sources. In this case,
very small or large are exponents less than E-32 or greater than E+32 .
There are other versions of spice (Contec for one) that handles these
exponents with no problem.
I would like to echo Arpad's comments that it really doesn't make sense
to utilize s-parameter characterization for the actual active circuitry.
As mentioned one would need to have s-parameter data at several different
points during the transition and then interpolate the data between sample
points and arrange for the right data to be used at the right voltage point.
Sounds quite messy and subject to error not mention probably being slow.
In addition, the transfer characteristic S21 and S12 would make sense
between input and output nodes, while the reflection coefficients S11 and
S22 would not really model the pull up/pull down characteristics. I also
have trouble seeing how the delay and slew rate information would be
carried through the model.
To summarize, I think s-parameters make sense for the passive
parts of a devices model when used with certain caveats in mind,
but don't feel that it really makes sense to use them for the non-linear
active sections and even if it did make sense, don't know of any
tools available to handle the problem in a clean fashion.