open-gpu-kernel-modules icon indicating copy to clipboard operation
open-gpu-kernel-modules copied to clipboard

Driver with open kernel module uses more VRAM than with closed module for same task

Open XZVB12 opened this issue 3 months ago • 8 comments

NVIDIA Open GPU Kernel Modules Version

580.82.09

Please confirm this issue does not happen with the proprietary driver (of the same version). This issue tracker is only for bugs specific to the open kernel driver.

  • [x] I confirm that this does not happen with the proprietary driver package.

Operating System and Version

Manjaro /Arch Linux

Kernel Release

6.16.8

Please confirm you are running a stable release kernel (e.g. not a -rc). We do not accept bug reports for unreleased kernels.

  • [x] I am running on a stable kernel release.

Hardware: GPU

NVIDIA GeForce RTX 3060

Describe the bug

More video memory is used when using the open kernel module than when using the closed kernel module. The difference is ~ 300-400 Mib when compared through nvtop or 100+ Mib when compared through nvidia-smi. This can also be tested by comparing how much video memory is available, for example when using closed kernel modules I can additionally to the work task start the hardware decoding video and everything will work, whereas when using open kernel modules video memory remaining during the work task not enough to run hardware video decoding.

I compared both kernel modules in the same scenarios: llama-server + 2 firefox browser. Computation task and firefox open tabs are completely the same. In the first case I have enough video memory to decode the video, in the second case I do not.

Image

Image

Image

Image

To Reproduce

1)Run a task that requires a lot of video memory 2) Compare VRAM consumption between open and close modules

Bug Incidence

Always

nvidia-bug-report.log.gz

nvidia-bug-report.log

More Info

No response

XZVB12 avatar Sep 24 '25 15:09 XZVB12

Is there any chance of getting an answer?

XZVB12 avatar Oct 14 '25 15:10 XZVB12

When you run the closed source driver, are you disabling GSP as well?

Binary-Eater avatar Oct 15 '25 15:10 Binary-Eater

When you run the closed source driver, are you disabling GSP as well?

yes, nvidia.NVreg_EnableGpuFirmware=0 in grub config

i also add nvidia module in mkinitcpio.conf for early load, if it important.

MODULES=(crc32c nvidia nvidia_modeset nvidia_uvm nvidia_drm)

XZVB12 avatar Oct 15 '25 15:10 XZVB12

@XZVB12 is the memory increase you see when GSP enabled just a static increase of around 300-400 MB?

Binary-Eater avatar Oct 26 '25 01:10 Binary-Eater

@XZVB12 is the memory increase you see when GSP enabled just a static increase of around 300-400 MB?

Yes.

I also use variables for hardware video decoding in firefox, maybe it also matters.


LIBVA_DRIVER_NAME=nvidia
NVD_BACKEND=direct

in /etc/environment

I have re-reviewed nvtop screenshots and the only difference in memory consumption that I see here is related to firefox , although I tried to maximize repeat usage scenario with number of open tabs, besides I checked it several times this is not an coincidence. I don’t think it’s a problem in firefox, because the only thing that changes is the Nvidia driver As far as I remember the difference in using video memory is noticeable and without firefox.

XZVB12 avatar Oct 26 '25 10:10 XZVB12

Unfortunately, the fixed size VRAM usage when enabling GSP is due to GSP needing to be loaded into the GPU's discrete memory, consuming 300-400MB. This is a larger problem we are actively looking into how to improve.

Binary-Eater avatar Nov 01 '25 07:11 Binary-Eater

Unfortunately, the fixed size VRAM usage when enabling GSP is due to GSP needing to be loaded into the GPU's discrete memory, consuming 300-400MB. This is a larger problem we are actively looking into how to improve.

Is there a chance that this will be fixed? 400 Mib considerable size and no desire to waste it. this is the reason why I don’t use open-gpu-kernel-modules

XZVB12 avatar Nov 02 '25 17:11 XZVB12

I cannot make any short term claims unfortunately. What I will say is we do recognize this as a serious problem and are spending effort thinking about how to resolve this internally.

Binary-Eater avatar Nov 02 '25 17:11 Binary-Eater