On Fri, May 30, 2008 at 10:06 AM, al davis <
ad151@xxxxxxxxxxxxxxxx> wrote:
On Thursday 29 May 2008, Dan McMahill wrote:
> al davis wrote:
> > Those fancy models that you get by downloading or from
> > databases are usually overkill. If you don't understand
> > them, use a simpler model.
>
> I think there can be some use for those vendor models, but
> you really have to watch out because it is not always clear
> what is modeled and what isn't. The result is you end up
> having to trace out the model or build up a model
> characterization test bench that compares some simulated
> values to the datasheet or real measured values.
That implies you know how to do that. Most people who can
really use the detailed models also know enough to make minor
hacks to them, like changing "table" to "pwl".
The problem is that lots of people just need a simple model, and
are misled into believing that the need something like that,
when in fact the detail could even be misleading.
If you are designing high performance analog IC's, you probably
need the detail. If you are using stock op-amps for audio
filters, I can almost guarantee you don't.
That is one of the principles of the "fast-spice" simulators.
They get their speed partly by using simple models, and partly
by algorithm tricks. They get away with the simple models
because they know the extra accuracy is not needed.
That is one of the advantages of gnucap. If you use the simple
models, and set the tolerances for the accuracy you actually
get with the "fast" simulators, it should run as fast as they
do.
On the other hand, if you use the detailed models and set the
tolerance tight, gnucap can be more accurate than Spice.
Unfortunately, with "more accurate" comes "slower".
> And I agree with Al's comment that they're often times
> overkill. In one case I can think of, all you really needed
> was the dominant pole to be modeled and you instantly could
> predict some extremely important behaviour. I can see
> wanting to look at distortion in an audio filter or something
> but then I have doubts about the models.
The time-step control in Spice (Berkeley Spice or NGspice) is
not accurate enough for meaningful distortion measurements
regardless of the models you use or tolerance settings. Gnucap
is if you tighten the tolerances and use good models. The
default tolerances are not tight enough for that.
Same goes for transient noise analysis.