Running on mixed devices simultaneously?

Accelerated Computing CUDA CUDA Programming and Performance
1

I was under the impression that you could not use a (for instance) 1.0 and 1.3 card simultaneously in the same app. I am not quite sure where I got that from though…

I did just try this on one of my systems - splitting a single computation across a C1060 and a Quadro FX 5600. The results appear to be correct, and the computation took less time than simply using either card.

Am I just mistaken in what I remembered, or is there some reason I should not do this?

Thanks!

2

At one point that statement was in the programming guide. I don’t know if it is is still in the version 2.2 guide or not.

All reports on the forums are that it just works. And the programming guide’s statement has been brought into question. I don’t recall any official word on it, though.

3

I can’t imagine that we’ll break it. There are potentially multi-GPU things that won’t work across generations (the multi-GPU GL interop on Quadro comes to mind), but that will be clearly described.

4

Thanks to both for your responses. This makes sense, and I’m glad that I’m not remembering something that didn’t exist. =)