CUDA on SLI connected cards No difference observed

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I’m testing CUDA on multiple graphic cards which are connected using a SLI ( Scalable link interface) cable so that they act as one. When I run the deviceQuery program from the SDK in both the cases I’m getting the same number of cores present. Shouldn’t we get double the amount of cores and multiprocessors from the deviceQuery program if we have two SLI connected graphic cards ?

Can any one please explain how to make this work?

Thanks in advance…

2

Hi Kartik,

If you have connected your graphics cards through SLI they will be treated as one, and you will not be able to extracted the power of both GPU’s using cuda. Since only one of them will be used. So yo need to turn-off SLI first.

This is explained the NVIDIA CUDA Programming Guide

You will be able to find it under the section “Multiple Devices” (Section 3.4 in ver 1.1)

So once you disabled SLI, Run the deviceQuery program and you will find, properties for 2 cards getting listed.

Also don’t forgot have a look at “multigpu” sample in the SDK.

3

Hi,

Thanks for clarifying that for me…

Cheers,
Kartik