GPU A40 passthrough poor performance

Hello everyone,

I am facing some issues with my A40 GPU while gaming. Currently, I have set up a vGPU with flavor A40-8Q, but the performance is quite poor. It appears that the FPS is capped. I’ve tried changing the vGPU flavor, but it did not improve the performance. I’ve also used ‘frame_rate_limiter=0’ to uncap the frame rate, but it did not help.

To address this, I decided to pass the entire GPU to a Windows 11 VM. However, I’m encountering problems while installing the driver. The installation process does not complete, and after a forced reboot, the VM fails to boot up.

At this point, I’m not sure what to do to either enhance the vGPU’s performance or successfully pass the GPU to the VM.

If anyone has any suggestions, I would greatly appreciate them. Thank you.

which Hypervisor is in use? FRL off should do the trick. I guess something didn’t work with the FRL off otherwise you should see >60fps. Please simply try to run a benchmark like Unigene Heaven to check if the GPU renders more that 60 fps.



Thank you for your reply. It sent me down a new rabbithole of testing.
I am using Ubuntu 22.04, qemu:7.0, nvidia: 510.108.03.

After running the benchmarks like you suggested, I saw an increase of almost 2x fps.
Exactly the result I was expecting. I’ve updated qemu and nvidia driver between the tests. This could have had the desired effect.

I’ve been playing around with different schedulers.
Could you tell me what is the difference between EQUAL_SHARE and FIXED_SHARE schedulers?
I see a drop to ~20fps during benchmarks. I’ve used both the default values and 1ms time slices.

Also is there a way to slice cuda cores to vGPU. I mean just like vram, is there a way to have a dedicated number of cuda cores for each vGPU?