Having discovered the issue with using remote desktop to run CUDA, I though I could get around it by logging into the machine with the target GPU (let’s call is GPUServer) using SSH to a daemon run by Cygwin. I managed to get the environment variables set up in bash for cl.exe and that all runs fine. But, when I run the built executable, I get the following error:
CUDA Error: no CUDA-capable device is available (at initialisations)
What is interesting is that it can find the card OK and recognise that it is capable of running CUDA, but it looks like the card is already locked.
I am logging in as the same user as GPUServer is currently logged in as locally. It is also interesting (to me at least) that if I ssh to GPUServer from GPUServer itself, I get the same error message.
I obviously am missing something in the way that Cygwin/SSH works.
If anyone can help me avoid having to resort to VNC, then that would be much appreciated.