GTX SLI or Tesla

Hi guys,

I’m wondering witch one is better for CUDA usage. Use two or more SLI cards in parallel or use one or two Telsa cards. There is a benchmark or comparation in somewhere?

It depends on what you want to do. For raw memory bandwidth and single precision floating point, the GTX 580 is the fastest card out there. If you want fast double precision, the possibility of 6GB of device memory, ECC memory protection, or bidirectional PCI-Express transfers the Tesla cards win. (There are other minor differences, but those are the big ones.) But keep in mind that If you don’t use double precision, you will find that Tesla cards are typically slower.

I should also note that CUDA does not use SLI for anything. You program multiple CUDA devices the same regardless of type (GeForce or Tesla, with or without SLI): create a CUDA context for each device, often in different CPU threads, and start kernels in each context. If you need to exchange data or synchronize execution between cards, you have to manage that yourself.

Hi, the double that your’re talking about is the C type double right? If yes I will need in maximum float, in the most cases integer variables.

Sorry for the ignorance :)

Correct. The C double type is a 64-bit double precision floating point number, and the float type is a 32-bit single precision floating point number.