FAH-GPU2 beta client is available!

Oh I didn't mean to suggest that. But I remember back when I was using two X1900's, Crossfire had to be disabled to get two separate clients (one for each card) to work.

 

I didn't know this can be unreliable

At least on a Nvidia 2 x 9800 GX2 system I've been able to fold with all four cores seemingly OK. The main problem being related only to the Vista OS, that is having to have all the VGA outputs enabled.

 

Could it be, that my CPU's L2-caches are too small for folding (2x256 kB)?

I've changed a little bit the memory's timings:
3-4-4-7 (CL-RCD-RP-RAS):

Then 2,5-3-3-7:

And now with overclocked CPU (2x2,3):

Command rate is still 2T (all 4 memory banks are used).

 

Yes. Folding seems to love large caches, and AMD's historically small cache sizes have hurt their folding performance relative to their Intel counterparts.

Memory timings are almost irrelevant on AMD platforms, in my experience (for anything outside of synthetic memory tests, of course). Latencies are already so low with the IMC that a clock cycle here or there isn't going to matter.

 

I wonder why there's such a large CPU-overhead on CAL/Brook whereas it's almost nonexistant for CUDA-based clients. Any insights as to that?

 

Woohoo!

 

Looooocal share. I suppose it is that miraculous for F@H whilst most other GPGPU apps don't seem to gain that much of a percentage on it.

 

Nvidia > Ati

how many apps run on CAL ?. ati seem on its knee now. CUDA 2.0 is here, and whatever ati say about folding, the fact is , they can't handle the GPGPU match agasint nvidia. yet , at least ...

 

Any news on FAH-GPU2 + CAL on Linux, Mike?

 

Is it just available resources that stop the Linux effort from progressing in parallel with the Windows one? It's pretty obvious you have more users on the Windows side, so prioritising things for that platform seems like what you'd do if you simply didn't have the manpower to do both at once.

 

If i may offer, I believe the issue is one of thread scheduling. The science is calculated via simulation cores written for a *nix environment, at least in terms of the thread scheduling model used (MPI - Message Passing Interface) so that is why extensive investment is needed on *nix development as well.