Nvidia Driver Fails to use system ram when vram is full leading to crashes

[deven.boutin]

Nvidia Driver Fails to use system ram when vram is full leading to crashes or performance problems. This is to gaming and not Cuda or other similar workloads where memory sharing is properly supported on Linux (as far as i can tell).
This is a technology supported by all uefi compatible gpu drivers and hardware providers including Intel and AMD for the past ten years except Nvidia’s linux driver. It is even the case that Nvidia’s windows driver even supports GTT memory sharing and yet the Linux driver doesn’t.

This is not a new issue and has been reported many times before and for many years in a row.
Here is a non exhaustive list:

Non-existent shared VRAM on NVIDIA Linux drivers

The NVIDIA Linux driver doesn’t handle the VRAM sharing with the system RAM. What I’m exactly referring to is this: Windows 10 Task Manager in GPU section As can be seen, there is the “Dedicated GPU memory” and the “Shared GPU memory” that is actual system RAM but shared with the GPU, so in case the GPU runs out of VRAM, the system or game doesn’t simply crash or have a drop in FPS. So what is the problem then? The problem is that the “Shared GPU memory” doesn’t exist on the NVIDIA Linux dri…

DRM GTT/Shared memory broken in Nvidia Open drivers on Linux

NO SHARED MEMORY FOR YEARS [NVIDIA_UVM] - BASIC FEATURE

Add support for shared memory · NVIDIA open-gpu-kernel-modules · Discussion #618

https://forums.developer.nvidia.com/t/vram-allocation-issues/239678

This is very obviously a duplicate issue and I called your phone support line about it and was told to create a post in the forum in order to obtain support. So here i am.

[jonwright.24]

The sysmem fallback is a nightmare. Instead of fighting the driver, I started shrinking the models. LoRA Lens uses SVD math to keep the weights small enough to stay in the dedicated VRAM. No sharing needed, no crashes. It’s the only way to get stability on Linux/Windows right now.

[deven.boutin]

I didn’t mention AI or LLM’s what are you talking about? Of course it’s related, but it’s also a total non sequitur.

[nick.janetakis]

@deven.boutin

This is a technology supported by all uefi compatible gpu drivers and hardware providers including Intel and AMD for the past ten years except Nvidia’s linux driver. It is even the case that Nvidia’s windows driver even supports GTT memory sharing and yet the Linux driver doesn’t.

I can confirm on my GeForce 750 Ti (2 GB card) on Windows, everything was ok to run any workloads I wanted and it offloaded things to system memory seamlessly. It did this since 2014 when I put together this system.

On the same system but running Arch Linux (any Wayland DE) with the latest 580 drivers, I could barely open a few browsers and terminals before going out of GPU memory and things would crash or stop working normally. I included way more details in one of the threads you linked at Non-existent shared VRAM on NVIDIA Linux drivers .

I ended up spending almost 2 full weeks troubleshooting as much as I could, wrote an 8,500 word blog post, made videos, opened issues on GitHub which got a reply from NVIDIA, but no resolution.

All in all, since then I still use the same system but I switched to an AMD GPU (RX 480) and this problem went away entirely. I’m only posting this piece of information to confirm it very strongly appears to be a problem with NVIDIA drivers but not AMD.