Nvidia control panel option: Optimise for Compute Performance

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

The deep learning program, that I’ve written in pytorch with cuda10, works if the parameter, that can be found in Manage 3D setting in Nvidia Control Panel, “Optimise for Compute Performance” is set to Off. However, I would like to set this to “On” for boosting the performances, since the program uses only a small part of resources, for instance, the GPU cores never reach the maximum clock frequency.

The problem is when I set “Optimise for Compute Performance” to “On” the error that it gives to me is the following:

"RuntimeError: CUDA error: all CUDA-capable devices are busy or unavailable".

How can I solve it?

My GPU is a GTX960m.

2

It looks like when it is in optimize for computer performance, it locks the gpu for the OS, who is the first to ask and never lets go…

I guess you’d need a second (could be much weaker) GPU for the OS so this one is free. Often there is integrated graphics (set in BIOS to start the integrated graphics first). Of course don’t start anything else that might want the graphics card.

Since you say it already isn’t pushing the GPU, probably you don’t need to reserve the GPU?
I find its notable if GPU goes past 50% on most things.