I was contemplating the use of a GeForce GTX 295 card in Dell Precision workstation for some CUDA work. After looking up the specs on the various Dell Precision towers, none of them support the required ~290 watts maximum per slot (the closest is the the 7500 which supports up to 225 watts per).
I noticed that it appears the NVIDIA Quadro 5800 uses the same processor as the 285, and the 285 is the single proc version of the 295. So…can I just use two FX5800’s for my CUDA work? What are the gotchas here?
As far as programming goes, there should be no difference. The GTX 295 shows up as two separate CUDA devices, so plugging in two separate cards should be equivalent. The GTX 285 chip is in fact a little better (wider memory bus) than each half of the GTX 295. Plus, each GPU on the GTX 295 has to share one PCI-Express slot for communication, so if you are planning to run code which needs to transfer lots of data to both CUDA devices simultaneously, you’ll probably see better host-to-device bandwidth with two cards.
Just to check, these computers have four of the 6-pin PCI-Express power cables inside? Each card will require two of them. Also, are you sure that the PCI-Express slots are spaced properly to take two double-slot cards? That’s the only possible gotcha I can think of.
You should go with the FX5800, they have 4GB each! Or you could go the (slightly) cheaper way and use Tesla C1060’s, which are really the ones you should be using for CUDA work anyway ;-)
There should be no reason for you to buy 2 5800, you can always get one 5800 and one c1060, and also remeber if you don’t need 4 Gig of RAM a 285 will be faster (has faster clocks and memory)
The major power supply for GTX295 is from the 6-pin and 8-pin power cables, only about 50watt is needed from the PCI-e slot of the motherboard according to reports. So I don’t think you need to buy two FX5800 instead of GTX295 just for the power supply problem.