L2 cache question

[Longshot]

Does 512 of L2 cache offer better performance in LL testing over 256 L2 at the same processor speed?

[Prime95]

Yes

[Peter Nelson]

Originally Posted by Prime95

Yes

Am I right in thinking this can be generalised to larger sizes also?

1024 faster than 512

2048 faster than 1024.

And is 2048 the largest cache size currently optimised for?

So eg when machine with 4096 cache is available, the software current client does not yet take advantage and will use code/behaviour as in 2048?

[ColdFury]

Am I right in thinking this can be generalised to larger sizes also?

Everything else being equal, yes, but increasing the side of caches can cause other side-effects (like increasing cache latency) which may hurt performance. With Prime95 its generally a win though.

[Prime95]

Originally Posted by Peter Nelson

Am I right in thinking this can be generalised to larger sizes also?

Sort of. Of course more cache is better if all other things are equal (same CPU stepping and L2 cache latency).

However, 512K is near-ideal for all exponents we are currently testing. So going to 1M and 2M doesn't buy much extra speed. 256K is OK for current double-checks. 128K is real bad.

You might try looking at http://mersenne.org/bench.htm to see the effect of L2 cache on timings. At some FFT size in the distant future a 1M cache will be very important for getting good timings.

[Longshot]

Thanks guys. It is nice to be able to tap in to all the expertise here.