Dedicate one Nvidia card for CUDA and another Nvidia card for Xorg?

Graphics / Linux Linux
1

As the topic says, I have got a cheaper NVIDIA card that I would like to use for Xorg and an expensive NVIDIA card that I would like to use exclusively for CUDA computation. Thus far, whatever I tried, I was not able to prevent the expensive card to be used by Xorg as well, for example:
fuser -v /dev/nvidia[0-1]
shows Xorg, x11 and akonadi on both cards which is confirmed by nvidia-smi showing a small amount of memory (2MB) without doing anything on the expensive card.

Is there a way to achieve this?

EDIT: just noting that I do not care much about the 2MB but from some reason, the expensive card gets “stuck” at Xorg from time to time (showing near 100% utilization and fan running at 100% endlessly)

2

Please try
Option “AutoAddGPU” “false”
in a Serverflags Section and
Option “ProbeAllGpus” “false”
in the nvidia device section.

3

Many thanks for your reply, unfortunately, that did not help. I have been actually using “AutoAddGPU” “false” already and the second option did not change anything - there are still 2MB used by x11 and related processes on the CUDA card, fuser still reports kwin_x11, plasmashell etc. These are the relevant sections of my config:

Section “ServerFlags”
Option “AutoAddGPU” “off”
EndSection

Section “Device”
Identifier “Device0”
Driver “nvidia”
VendorName “NVIDIA Corporation”
BoardName “GeForce GTX 1650”
BusID “PCI:66:0:0”
Screen 0
Option “RegistryDwords” “RMUseSwI2c=0x01; RMI2cSpeed=100”
Option “PrimaryGPU” “on”
Option “ProbeAllGpus” “false”
EndSection

4

Please run nvidia-bug-report.sh as root and attach the resulting nvidia-bug-report.log.gz file to your post.

5

Attached. EDIT: attachment removed

6

Xorg is only running on the 1650, so the config options are fine. You should configure and start nvidia-persistenced though, to keep the 1080 alive.

7

Oh, fine then, many thanks. I thought that fuser showing kwin_x11 etc on both nvidia0 and nvidia1 means that both cards are used by it… but I can see a difference of mmaped mode and I am not sure whether this was the case the last time my X was frozen by the 1080 misbehaving (which I want to prevent). So for the moment I’ll just see whether it’s indeed stable now and look at your suggestion of nvidia-persistenced.

8

Of course, also set CUDA_VISIBLE_DEVICE to not interfere with xorg.

9

If anyone with the same problem reads this, it turns out that none of the suggestions thus far solved the problem. It’s still not completely dedicated - Xorg is still somehow connected to both cards and if an issue happens, it’s still impossible to reset the CUDA card without Xorg restart.

I have opened a new thread about it (thus far without a solution but if anything is discovered, I’ll write it there):