Re: [ng-spice-devel] ACS
---"Alan" == Alan Gillespie <alan.gillespie@analog.com> writes:
Alan> But my understanding is that spice only iterates until
Alan> things stop moving within reltol, so surely the final
Alan> solution will never be anywhere near accurate enough to
Alan> need 64 bits of precision ? So long as the matrix code
Alan> produces results accurate to within a few bits at single
Alan> precision (is that about 24 bits, i.e. 1 in 16 million)
Alan> that'll be way more accurate than the 1 in a thousand
Alan> or so that reltol is usually set to.
Alan> Am I missing something ?
The nonlinear convergence criterion is a separate issue from the
accuracy of the linear solver. Under most circumstances, the nonlinear
solver tolerance is loose enough that even if the linear solver is
inaccurate, things will still converge. But, given a really bad
condition number for the matrix, disastrous cancellation or other
things that can happen, the nonlinear solver won't converge because it
is being given a trashy solution by the linear solver. Or if it does
converge, it takes more iterations.
I've also seen situations where a reltol of 1e-8 is warranted -- for
instance, where someone is trying to get the settling time of an
amplifier down to microvolts on a 1 volt signal. This starts to get
interesting, since the voltages must have more than 8 good decimal
digits, which is sometimes on the edge of what the simulator can
resolve.
--Steve
Partial thread listing:
- Re: [ng-spice-devel] ACS, (continued)
Al Davis
ACS converted to automake, autoconf
Arno W. Peters