How can I use a discrete GPU only for training?

Hi all,

Is it possible to use a NVIDIA GPU card only for ML training ?

I have a problem with the NVIDIA driver which I have mentioned here:
Blank screen after installing nvidia-driver-510 on RTX 2060 SUPER machine - Graphics / Linux / Linux - NVIDIA Developer Forums

Now when I select only IGFX in BIOS and install Ubuntu, the iGPU is used properly and there are no issues. In all other cases the dGPU is prioritized for display. And the nvidia driver has problems still not solved.

The question is whether I can install an nvidia driver which runs only for providing the CUDA service (for ML training) while all graphics runs on on iGPU ?

I read on the nvidia-prime selection tool, but this also assumes that the driver is working properly. I also see that there are are some server drivers, but not sure whether its useful or not.


Hi @rreddy78 ,

From the looks of the other thread it seems you go the driver issue resolved?

Just for completeness, you will need to have a working driver on your system to use CUDA for ML. Regardless of wether you use the GPU to display anything or not.

I hope this helps.

Yes. My issue is resolved.
OK. Thanks. I will have to test this configuration and see.