Re: [ng-spice-devel] ACS
Steve,
It was, indeed, SuperLU I was thinking of. I assumed
it would be faster as a straight swap, so it's interesting
to know that you didn't find that.
However, I was also interested for a couple of other
reasons -
Firstly, I've noticed a significant difference in the
speed of Spice3 for large circuits between the old Intel
processors with 512K cache versus the Celeron with only
128K cache. It seems that the new processors are all 256K
cache, so I suspect they'll pay a penalty when it comes
to large Spice3 simulations. Did I read that even the
Athlon was going down to 256K cache ? SuperLU claims to
be particularly suited to machines with limited caches.
Secondly, they mention something about being able to
reduce the numerical error by iterating after the direct
solve. I was wondering if using single precision for the
direct solution, and optimising it for the new 3D-Now/SSIMD
instrucions, and doing a few iterations afterwards, would
lead to a faster overall solution. (3D-Now and SSIMD don't
do double precision yet, as far as I'm aware.) Using 32 bit
variables would presumably allow bigger arrays into the
small cache, as well.
What size of cache and circuit did you test it with ?
Cheers,
Alan
Partial thread listing:
- Re: [ng-spice-devel] ACS, (continued)