Foldy CPU Vs. GPU times

diogenese · 23 Dec 2006 at 22:55

I'm running an X2 4400 with a X1900XTX at 675/800. I have been experimenting with different cpu usage on the second core, the first core is folding 100%
With no CPU folding I get a frame every 6 minutes exactly, at 50% cpu I'm down to 10 minutes per frame and at the moment I'm testing at 25%.
This doesn't seem to hold with Stanfords claim that " it's not actually using any cpu power" theory.
I was only trying to clear a half finished before it times out but it's got me all confused as to how much cpu power the gpu client is really using?

joeyjojo · 23 Dec 2006 at 23:07

Thought it took a whole core :confused:

rich99million · 23 Dec 2006 at 23:24

Taken from the FAQ

Client looks like it is using lots of CPU time: Graphics drivers must poll the GPU to see if it is done. This will look like a lot of CPU time being used, but nothing is really being done, so one can run other relatively CPU heavy code at the same time. However, we do not recommend running multiple FAH clients, as this can significantly slow down the GPU client. A good rule of thumb: leave at least 25% of the CPU time available for the GPU client.

well that is quite simply utter nonsense - as you have found for yourself

I know when it was just released they were confident that you could still use most of the core and the GPU wouldn't be affected, I didn't realise they were still following a similar line in the FAQ though :confused:

joeyjojo · 23 Dec 2006 at 23:28

So what would you recommend? Set FAH to 75%? 25%? Off?

Edit: I think this is my 1337 th post. I missed it at the time

diogenese · 23 Dec 2006 at 23:34

I've run a couple more tests, nothing extensive but here it is.

00% cpu, 6mins/frame
25% cpu, 8mins/frame
50% cpu, 10mins/frame
100% cpu, 25mins/frame

Looks like you can't fold on the cpu at all for maximum returns :confused:

Mattus · 24 Dec 2006 at 00:16

What I don't understand is that you can run a GPU on a much slower CPU and get similar results. If 100% of, say, an Athlon 1100 is enough, then why is 100% of a much more powerful CPU still needed?

diogenese · 24 Dec 2006 at 00:27

I'd love to know too, it confuses the hell out of me!

rich99million · 24 Dec 2006 at 00:58

ahhh just found clarification of something I thought I'd seen around before...

P5-133XL said:
So far, the best of that type of technique is using a HT P4. Then you can CPU fold and then let the virtual CPU service the GPU. That seems to work well with the CPU netting around 50% while the GPU only drops only 2-5% producing a net total gain.

It comes down to the GPU client needs a full CPU to service it. It doesn't seem to matter as to the capability of that CPU, but it does seem to matter that it gets 100% of that CPU.

So an AGP card on an HT P4 could be a great value for money upgrade as you only lose the 20% or so increase in production by using the "virtual CPU" for CPU folding but you gain the whole of the GPU output which more than makes up for it :cool:

Just a shame that old P4s are quite inefficient - also I don't have one

SiriusB · 24 Dec 2006 at 01:05

Any DX9 experts here at all? I really am curious why CPU speed has little or no affect on GPU folding but the amount of CPU time does.

SiriusB

diogenese · 26 Dec 2006 at 08:15

I've been running Muon1 overnight, it can make use of multiple cores so it has been at 100% of both cores whilst my gpu has been snatching the odd 2-6% cpu time. now this is odd, folding frame times are now 14 minutes :confused:

So doing other work on the cpu does not give nearly the same performance hit as running another instance of folding, is this a deliberate thing to stop you running multiple instances :eek: