Mixing cards (shaders)...Rules?

[OldChap]

Ive read, and pointed others at, threads about performance reductions when mixing the number of shaders on 9 (and 8) series cards.

Can anyone tell me if this problem still exists when using 200 series cards

[alexopth69]

yes it does. shaders number must be an integer fraction on eachother. So I have used GTX285 and GTX260 and 260 produced a poor 3000 PPD

[OldChap]

Thanks, I thought this would be the case but my poor memory was recalling a comment otherwise

[Leonardo]

shaders number must be an integer fraction on eachother

I know that a GPU with 128 shaders - one GPU of a 9800GX2 does not mix well with a GPU of 112 shaders - 88/9800GT. I don't understand your mathematics statement. Would you please illustrate that?

[P5-133XL]

128/112= 1.1428 (not an intger factor)
112/88 = 1.2727 (not an integer factor)
112/56 = 2 (an integer factor)

[chungenhung]

128/112= 1.1428 (not an intger factor)
112/88 = 1.2727 (not an integer factor)
112/56 = 2 (an integer factor)

That is some new information. Never knew that before.

[codysluder]

I don't think anybody knew it before now. I certainly don't remember reading that until today.

[Xaero]

Okay, so heres the question and I probably look dumb.
Say I have a GTX 280, and I got a 8800 laying around...
I know you can't SLi them... but could I leave the 8800 headless plugged into the same motherboard as the 280 and be able to fold on it?
I've been reading about it but want to clarify it before I have my old 8800 shipped halfway across the US...
thanks
~Xaero

[jaak ennuste]

Okay, so heres the question and I probably look dumb.
Say I have a GTX 280, and I got a 8800 laying around...
I know you can't SLi them... but could I leave the 8800 headless plugged into the same motherboard as the 280 and be able to fold on it?
I've been reading about it but want to clarify it before I have my old 8800 shipped halfway across the US...
thanks
~Xaero

You can use many different cards. They have separate clients with separate work directories. SLI and Physixs should be disabled, use dummy VGA plugs to fool monitors and extend desktop to each card in windows display properties. Clients should have different machine ID-s (folding@home -> setup -> advanced) to avoid conflict.

I have run 2 GPU GTX 295 and 1 GPU 280 same time, no problem whatsoever.

cheers

Jaak

[codysluder]

You can use many different cards. They have separate clients with separate work directories. SLI and Physixs should be disabled, use dummy VGA plugs to fool monitors and extend desktop to each card in windows display properties. Clients should have different machine ID-s (folding@home -> setup -> advanced) to avoid conflict.

I have run 2 GPU GTX 295 and 1 GPU 280 same time, no problem whatsoever.

Multiple cards will work, but not necessarily effectively. If you mix multiple GPUs of different performance classes, Windows will still send updates to all parts of the "screen" in a synchronized process so you won't get full performance out of both GPUs.

[dempaSD]

I have run 2 GPU GTX 295 and 1 GPU 280 same time, no problem whatsoever.

Jaak,

GTX 295 and GTX 280 works fine without any performance decrease since those use the same amount of shaders for the GPU's.

[jaak ennuste]

You can use many different cards. They have separate clients with separate work directories. SLI and Physixs should be disabled, use dummy VGA plugs to fool monitors and extend desktop to each card in windows display properties. Clients should have different machine ID-s (folding@home -> setup -> advanced) to avoid conflict.

I have run 2 GPU GTX 295 and 1 GPU 280 same time, no problem whatsoever.

Multiple cards will work, but not necessarily effectively. If you mix multiple GPUs of different performance classes, Windows will still send updates to all parts of the "screen" in a synchronized process so you won't get full performance out of both GPUs.

I don't buy that. There is nothing to do with windows anymore. CUDA interface opens direct access to streaming processors, by-passing Windows graphics engine and every GPU works at full power being fed directly from CPu and main RAM. Very good introductory article is here: http://en.wikipedia.org/wiki/CUDA

Jaak

[dempaSD]

Ok, I have a spare 8800gs lying around. Will pair it with two 9800gx2's this week to verify for myself how it works. I know the expected "speed" for all of these running paired with only same type of GPU's.

[Hgradio1]

My experience says just dont try to mix 200 series with 88 or 98 series because they normally don't work at all with many systems.

[dempaSD]

Tried it:
Added one 8800gs (96 shaders) together with two 9800gx2 (128 shaders). Total five gpu clients. The 96 shader card produces ~25% less PPD like this than when paired with another 96 shader card. Really. And when not calculating PPD but rather the number of shaders, 96/128 = 0.75 ...voila.

I'll keep the 88gs folding overnight tigthly stuck between the 98gx2's to measure stability. (gpu0+1,gpu2, gpu3+4) I don't have other box to put the 88gs in anyway. Overclocked as far as cards allow in my not optimally cooled setup.

There, same experience for me as for most other people I've seen posting results here before.