To SLI or not to SLI?

[ET_]

If you have 2 GPUs on your system, should you run 2 instances of mfaktc (one for each GPU) or 1 instance with the boards set in SLI?

Guess it's been already asked, but I forgot the answer...

maybe worth to put it in the "GPU for dummies" document?

Luigi

[TheJudger]

Hello Luigi,

AFAIK you can't use SLI for CUDA, you'll allways have multiple GPUs do deal with, not a virtuall big GPU.

Oliver

[bcp19]

Originally Posted by ET_

If you have 2 GPUs on your system, should you run 2 instances of mfaktc (one for each GPU) or 1 instance with the boards set in SLI?

Guess it's been already asked, but I forgot the answer...

maybe worth to put it in the "GPU for dummies" document?

Luigi

Also, depending on the GPUs you will probably need 2 instances of mfaktc/o just to max each GPU.

[LaurV]

I have 2 GPU's and I DO use them in SLI, as the computer is much faster for OTHER activities in this configuration (AutoCad, SolidWorks, Altium Protel, etc). This has nothing to do with CudaLucas or Mfaktc, which are "small" things, and working in double precision. SLI addresses BIG things, like matrix multiplications, 2D/3D transformations (rotations of millions of 3D points in the same time, for example), mesh/2D/3D rendering and other stuff, and generally SINGLE precision, very-very-very fast single precision stuff. See the video demos on nVidia site, and the FAQs. SLI transmits "syncho, pixels and display data" over the SLI connection, therefore taking the burden out from the PCIe slot. Single precision calculus for video data can be scaled and split between few cards, therefore it is executed few times faster in parallel in all cards (SLI antialiasing, rendering, etc). CudaLucas uses double precision and it is not really aware of your SLI. It will still use ONE GPU only, always the first one, no matter how many copies you run, all of them will fight over the resources of your first GPU, the other one will stay empty even if you have SLI, and you will need the -D switch to force CudaLucas to run on the second (or third, etc) card. And mfaktc is even smaller, you can not max one GPU with only one instance of it. You need more instances to maximize one, and you will need two times more to maximize two GPU's, so it is not exactly the thing that would be aware (and extend itself under all of the blanket) of your SLI. You still can execute 100 copies of mfaktc on the same GPU and keep the second one empty, to be used for something else. No comment about the lousy performance you will get, this is just as an example. They communicate over PCIe, not over SLI wires.

With or without SLI makes no difference for CL and mfaktc on my computer. But using SLI is better for other bigger stuff (mentioned above) which IT IS "SLI-aware". Try fold-it for example, as an example of DC (distributed computing) game-like stuff, that take advantages of SLI, and see the difference when you wiggle and shake that molecules with or without SLI. There is also no difference in running CL or mfaktc if I set the physx on one gpu or on the other, or even set physx on mainboard, which has its own intel video card with intel GPU (unfortunately it can't run cuda :P) and let the two GPU's free, CL and mfaktc still run at the same speed.

[Dubslow]

Originally Posted by LaurV

And mfaktc is even smaller, you can not max one GPU with only one instance of it. You need more instances to maximize one, and you will need two times more to maximize two GPU's, so it is not exactly the thing that would be aware (and extend itself under all of the blanket) of your SLI.

Actually, one instance puts my card at about 80% load, and it's not worth the extra CPU core to add another.