Subject: Re: Number of Significant Digits
From: AC (comeral@ix.netcom.com)
Date: Thu Feb 28 2002 - 11:55:53 PST
"With more digits, you need more data points. With more
data points, you need more significant digits...."
I don't follow the correlation indicated above, at least without some
more
explanation. I always thought that the more significant digits you used,
the better,
as long as the additional digits represent a closer approximation to the
actual datapoint (not just noise). (Numerical or other noise which does
not represent actual physics being modeled does need to be cleaned up
when the model is created).
After all, when you use N significant digits, you're just saying that
the actual value has 0's in all succeeding decimal places, to the
accuracy of your approximation.
The problem I've had when building models with tool-provided data
(SPICE) of limited precision is that when the tool truncates or rounds
successive values, it makes them either closer to the same value than
they really were, or further apart, adding an additional source of
numerical noise for no good reason. Sometimes the data says the values
were identical (e.g. to 4 digits) when this is clearly not the case, so
the model builder gets to decide which of two identical points to
include and which to throw out. In this case, I think the more digits,
the better, since no additional noise source is added. If you used SPICE
data with 30 digits of precision, for example, (or whatever
quad-precision works out to be equivalent to) to build your IBIS model,
then it would be equivalent to what SPICE would hand off directly to
your IBIS simulator without your intervention. Even without increasing
the number of datapoints at all, it should be a better approximation
than with an intervening step of reducing the accuracy.
Of course, if the model is based on measurement data, there may be lots
of digits of precision, of which many are not actual data but only
represent noise in the measurement. Only the engineer taking the
measurement can determine how many significant digits of precision they
have in the dataset. In my experience, it still may be necessary to
filter out noisy datapoints.
In either case, however, I don't see the direct correlation between the
number of datapoints needed in a model and the number of digits of
precision used to represent each point.
Alan Comer
Al Davis wrote:
>
> On Wednesday 27 February 2002 09:57 am, Lewis, Tony L wrote:
> > I am wondering if there are any rules or common methods for
> > deciding how many significant digits should be used when creating
> > an IBIS model?
>
> If you are asking this question, you should also ask about guidelines
> for determining how to sample a waveform or I/V table. They are
> related. With more digits, you need more data points. With more
> data points, you need more significant digits.
>
> The key to picking the proper resolution (in both) is to look at the
> derivative. You should use enough digits, and the appropriate
> samples, that the derivative looks good. This is more digits than
> intuition would suggest, and fewer samples than intuition suggests.
> You might need to smooth the data to get a reasonable derivative.
>
> Even though the derivative is not explicitly specified, it is
> actually more important than than the values themselves. In
> addition, it tends to magnify errors and noise.
>
> You should never have more than 2 adjacent points with the same
> value. If you get this, you should drop some points or use more
> sigificant digits so they are not identical. If you use more digits,
> make sure they are truly sigificant, not just noise. Even 2 adjacent
> points with the same value are usually wrong, unless you really mean
> to say the derivative is zero.
>
> Speaking of dropping points ... If the derivative is the same on
> successive points, you should drop some.
>
> For a I/V table, the derivative is di/dv, or incremental admittance.
> The reciprocal is incremental resistance. This is the value that
> determines how transmission lines are terminated, so it must be
> correct if you want correct modeling of reflections.
>
> For a wave table, the derivative is dv/dt, which enters into the
> calculation of current with a capacitive load. Errors in the
> derivative cause large errors in load current with reactive loads.
>
> Considering the importance of the derivative, I find it strange that
> none of the tools address it. There are tools that show you the
> waveform on a scope. It is easy to also show the derivative, which
> would be very helpful at revealing model problems.
This archive was generated by hypermail 2b28 : Thu Feb 28 2002 - 12:13:28 PST