Super-Newbie question Running CUDA apps on my machine

This is a truly noob question… I’m anticipating a collective :sigh:…

I’m expecting delivery of my second Geforce 8800 GTX tomorrow. I primarily intend to use it in SLI for gaming fun :w00t: , however I’m also a developer by trade and have a huge number of things I’d like to look at using CUDA for, having already played around with extending algorithms into SSE :geek:.

Reading through the release notes this is probably pretty clear, however I just need someone to take me exactly through the process I’ll need to do to switch between “games playing” mode and development mode.

My understanding is that I must not run any CUDA app on a card that has a display associated with it at all. To me, this suggests that I must disable SLI mode because presumably with it on, Windows (for that will be my development platform. Yes, I know it’s ugly but it’s the only environment I know!!!) treats both as being attached to the display.

Is this correct?

Finally, must I always be running the CUDA release of the display driver to avoid having to keep uninstalling and reinstalling drivers to be able to switch?

I understand that the future version of CUDA 1.1 has integrated the runtime into the driver, so this won’t be as much of an issue.

Many thanks in advance for any response.

This is not true in the general cause. My machine has only a single GPU, and I run computation and display on the same card all the time. The only limitation comes when you want to run a single kernel invocation for more than 5 seconds. In my work so far, the longest kernels I’ve written only take 10’s of milliseconds to run so this is not a problem. Just don’t go moving windows around while benchmarking your code, that slows down the computation :)

If your CUDA apps use only a single GPU, you can leave SLI enabled just fine. However, if you want to write CUDA code and run it on multiple GPU’s, you must disable SLI. CUDA will not automagically make your code run faster with two GPUs, you must break up the tasks yourself: and disabling SLI makes 2 CUDA devices visible to your code.

My advice is to just go with CUDA 1.1 from the start. It is much better than 1.0 in many ways. But I will add that I had no problems running games on my CUDA box with the 1.0 CUDA drivers installed. I didn’t thoroughly test this though.

Hi there,
That’s fantastic advice - thanks very much - answers all my questions! :thumbup: