Multi-GPU Requirements ! using CUDA with multiple GPUs


I want to know the GPU requirements for using CUDA with multiple GPUs. I want to use 3 cards. I understand we should disable SLI mode in order to use multiple GPUs with CUDA, but Should they still be connected using a 3-way SLI connector if I want to program using CUDA. And also, should the cards be of the same type ? Please let me know the GPU hardware requirements of using 3 cards with CUDA.

You do not need to use th SLI link with CUDA at all. WDDM in Vista and Windows 7 probably will insist that all the GPUs are identical, Linux doesn’t have that limitation. After that it is all down to your programming model. CUDA doesn’t have any facilities to “automagically” do multi-gpu, you will need to manage all of that in your own code, or use something like gpuworker.

You don’t need to disable SLI to use multiple cards, this requirement is no longer valid since some recent driver version.
AFAIK SLI is now totally orthogonal to multi-GPU CUDA.

As for hardware requirements, any motherboard with enough PCI-E slots should do really. Of course there are better and worse. There are threads about choosing the mobo on this forum, do a search. Any CPU will do as well. It’s been said that it’s best if there’s one core per GPU but it’s not really a requirement.
You will need a PSU able to supply the GPUs with juice, pay attention to the current @ the +12V line(s) and the number and type of connectors (6-pin, 8-pin).

I don’t think the cards must be of the same type. Many use strong GPUs/Teslas to do heavy lifting and a low-power GPU for visualizing the output.

You need to factor in the SMPS for powering all the GPU cards.