Deployment requirements for non-developer machines? Colleague's computer responds, "non CUDA

What software needs to be installed on a non-development machine to make calls into the CUDA runtime API?

I’m developing with the CUDA Toolkit 4.0, (MAY 2011). [font=“Courier New”]cudaRuntimeGetVersion()[/font] returns [font=“Courier New”]4000[/font]. I x-copy deploy my executable, DLLs and [font=“Courier New”]cudart64_40_17.dll[/font]. It works great on my development machine and on my Amazon Cluster GPU instance. On my colleague’s computer, it fails at the first CUDA call, which happens to be [font=“Courier New”]cudaGetDeviceCount(int *count)[/font].

When my colleague first tried to run the program, he got this error:

CUDA driver version is insufficent for CUDA runtime version.

Therefore, he got the latest driver from

Now, the error is:

non-CUDA capable device is detected


He has two GTX 480s, which are certainly CUDA-capable.

I’m trying to avoid installing anything more than necessary, for when we roll this out to more users.

What do I need to install to get this running on non-development machines?


EDIT. I should have mentioned that all machines are Windows 7.

The second error seems strange. What you need is a current enough driver for the CUDA version you using and the cudart DLL. Did you restart the machine after you installed thee new driver. I is sometimes needed

Thank you, laughingrice. That was the (easy) fix!


Might I suggest that this is something that NVidia should document in a prominent place? I expect to be getting a beta version of our code to non-developers in the near future: it’d be nice if there was some sort of howto for getting appropriate driver versions &c. I really hate it when I have to say “Well, duh, it works on my machine” :-)