I'm straightening out a friend's new system with a Kuma cpu. I was
astonished to find that prime95's torture test doesn't load up a core
100% -- not even close.
Ed Light wrote:
> I'm straightening out a friend's new system with a Kuma cpu. I was
> astonished to find that prime95's torture test doesn't load up a core
> 100% -- not even close.
>
> Any ideas?
The program could do "small" or "large" FFTs. A large
one, won't fit in CPU cache, so there are lots of memory
accesses as a result. A small one, might fit within the CPU
cache, so there would be reduced memory activity. The small one
should run at 100%.
The reason for using large FFTs or blend, is so the
memory gets tested.
The reason for using small FFTs, is to heat up the
processor, for those cases that you want to verify
the cooling. There are likely programs better than
Prime95, for heating. So it isn't a record setter in
that respect.
In terms of memory settings, some AMD processors
have options like "linked" or "unlinked". "Unlinked"
allows dual channel memory, to be used as two separate
memory banks, and for multithreaded or multiple
program usage, it might be a better setting.
"Unlinked" may even be the default. You could
play with that, or generally with memory frequency
or timing, to try to do better than you are right now.
( Some mention of Kuma here. There is L3, so this is a more
recent architecture dual core.)
My Athlon 64 PC was stolen and I've built a P31/Pentium E5200, for the 3
PCI, 3 PCI-E X1, and PCI-E 16. Lots of slots. On it, I've found that you
should copy prime95 to another folder, start both prime 95's, and set
each to 1 of the cores using the p95 Affinity setting. Then it works ok.
Ed Light wrote:
> I'm back.
>
> Haven't played with the Kuma again yet.
>
> My Athlon 64 PC was stolen and I've built a P31/Pentium E5200, for the 3
> PCI, 3 PCI-E X1, and PCI-E 16. Lots of slots. On it, I've found that you
> should copy prime95 to another folder, start both prime 95's, and set
> each to 1 of the cores using the p95 Affinity setting. Then it works ok.
>
I haven't really seen a problem here, and using something like this
seemed to be OK.