SMP & GPU folding

[jrabb1920]

Hey guys I'm SMP folding with a q6600, and have a ATI X1900GT. can I fold with my GPU and still fold SMP???

[uncle fuzzy]

It's not supported/recommended. One of my teams mates did it for awhile with moderate success, but she's a super geek.

[jrabb1920]

Thanks for the reply, maybe they will do a fix to allow it in the future, Thanks.

[uncle fuzzy]

The SMP wants 4 cores. The GPU wants 1. That's 5 total. Giving any cpu access/time to the GPU will slow the SMP. The GPU will also run slower because it's wasting time fighting for the cpu. Both together will yield less than the total of each of them running seperately.

[bruce]

The SMP wants 4 cores. The GPU wants 1. That's 5 total. Giving any cpu access/time to the GPU will slow the SMP. The GPU will also run slower because it's wasting time fighting for the cpu. Both together will yield less than the total of each of them running seperately.

If you're running Vista (i.e.- DirectX 10) it's a lot more realistic than if you're running DirectX 9. The GPU client does best when it's allowed to hog 100% of a single CPU. With DirectX 10, the client only takes 10% -- 15% of a CPU, leaving most of it to other things (including running an unsupported SMP client). All of the unofficial reports that I've seen of people trying it were from WinXP (DX9) and if somebody tries it on Vista (DX10) I'd like to hear the report. In fact, since the results on a Dual-core CPU will be different than on a Quad, more than one report will be worth reading.

We do know that a new GPU client will be out "soon" so it's probably worth waiting to see what's in it, since any such reports may become obsolete.

[jrabb1920]

Thanks again, I think they should try and write a program that dosn't need CPU support. but I wait and see what the next version has to offer.

[jrweiss]

The SMP client may "want" 4 cores, but it is entirely happy with 2 fast ones. I don't know why it wouldn't work on 3, plus whatever's left over from the GPU I/O...

[uncle fuzzy]

It works on a dual core by running 2 of the 4 FAHcores on each physical cpu core. That's an easy process to program into the client. How is it going to distribute those same 4 FAHcores across 3 or 3 1/2 physical cores?

[bruce]

It works on a dual core by running 2 of the 4 FAHcores on each physical cpu core. That's an easy process to program into the client. How is it going to distribute those same 4 FAHcores across 3 or 3 1/2 physical cores?

When you divide up the processing of the WU into four roughly equal quarters, they're inevitably, to some degree, unequal. Whether the four FahCore_A1 processes are running on 2 or 3½ or 4 CPU-cores, there are regular synchronization steps which force the faster processes to wait until the slower processes catch up. It's not fundamentally any different if you ask your computer to run four FahCore_a1 processes plus surfing the 'net or displaying data on your screen or managing data for FAH-GPU. There's no inherent limitation that requires each FahCore to be locked into a particular CPU-core.

[uncle fuzzy]

It works on a dual core by running 2 of the 4 FAHcores on each physical cpu core. That's an easy process to program into the client. How is it going to distribute those same 4 FAHcores across 3 or 3 1/2 physical cores?

When you divide up the processing of the WU into four roughly equal quarters, they're inevitably, to some degree, unequal. Whether the four FahCore_A1 processes are running on 2 or 3½ or 4 CPU-cores, there are regular synchronization steps which force the faster processes to wait until the slower processes catch up. It's not fundamentally any different if you ask your computer to run four FahCore_a1 processes plus surfing the 'net or displaying data on your screen or managing data for FAH-GPU. There's no inherent limitation that requires each FahCore to be locked into a particular CPU-core.

In other words, just go ahead and run the SMP and GPU together.

[bruce]

In other words, just go ahead and run the SMP and GPU together.

No, that's not what I said.

From a technical standpoint, it can work. For the beneifit of science, the Pande Group recommends against it.