One eternity later, still a problem…
I switched to Linux because more and more open sourced AIs are easier to install on Linux, but if we can’t use RAM as shared memory we’ll need A100 GPUs very soon in our PCs…
One eternity later, still a problem…
I switched to Linux because more and more open sourced AIs are easier to install on Linux, but if we can’t use RAM as shared memory we’ll need A100 GPUs very soon in our PCs…
Till now this problems still exists, and since I owns far more ram than vram, that really limited my ability to train and run large models
I am absolutely disgusted and baffled this is still an issue and nothing is being done or being said about this absolutely major issue. For example, this makes games that run close or at 8gb vram (like Squad) completely unplayable. The moment the vram gets to the limit the system freezes.
I will never buy another overpriced garbage nvidia gpu. You are treating us customers like complete crap.
Meanwhile you keep pushing your limited 8gb Vram cards, at the same time you neglect your growing linux consumer base. Completely unacceptable.
I mean is this even active forum? Is there some other place this issue should be brought up in? Because it seems that no one from Nvidia is seeing this post…
I just posted a workaround here:
https://forums.developer.nvidia.com/t/vram-allocation-issues/239678/57?u=majteam
Good game to all.
As what I posted there, this is still only a specific solution for dxvk (although thank you providing that solution!)
The problem still exists…
I’m commenting here to make sure that nvidia doesn’t see this problem as SOLVED.
This is not working in squad. Also I read that this limit is “guideline only” and game can just ignore it.
-Squad in dx11 mode.
-dxvk.conf in binary folder next to squadgame.exe
-In the file reads:
dxgi.maxDeviceMemory = 6144
Mangohud confirms nothing changes.
Hard NOT SOLVED
Edit: I also added this to the conf: dxvk.hud = memory
That does give me the hud ingame so I can confirm the file is in correct place and is being read, but max memory command is just being simply ignored.
Please reply something this thread is 1.5 years old…
It’s not a hard limit, games can still decide not to use it.
Also I think it is important to note that this issue is not only gaming related. It is also productivity issue. As someone said here already, this is a problem with their LLM’s. For me, I have encountered “cuda out of memory” error in blender, and for that only fix was to install windows alongside just to render the animation in windows instead of cpu in linux… That is a huge no no.
Nvidia, please fix.
I made an update to the workaround:
https://forums.developer.nvidia.com/t/vram-allocation-issues/239678/60?u=majteam
As this command “dxvk.hud = memory” does work when the config is in the game exe folder, doesn’t that mean that the config file is properly loaded and the dxgi.maxDeviceMemory command is just ignored?
Yes, the command seems to be ignored and although I don’t understand why, i don’t know the game SQUAD.
Unfortunately, some apps may choose to ignore this command. I have no other solution to offer you at the moment.
Edit:
After a brief search, your game seems to be particularly VRAM hungry for textures. There seems to be an UNCAP TEXTURE POOLSIZE option in the game settings. Could this possibly be the cause of your problem?
Of course not, textures low and this setting is not turned on.
The texture use is not a problem in windows 11. Because the nvidia drivers handle vram, unlike in linux. That is not a solution, this thread needs someone from nvidia to acknowledge this.
This is not only a gaming issue as i said.
0 answer from the devs EVER on this subject is crazy.
If it’s not actually non-existent, then there’s something seriously wrong with VRAM spillover performance vs Windows, because I can see that when the VRAM fills up, the system RAM increases, so it is actually using system RAM like Windows, but the application slows to an almost unusable state; meanwhile, in Windows, the application still runs normally.
Oh, and even though it’s actually using system RAM, somehow you are still unable to launch other applications while the VRAM is full; they just instantly crash because of full VRAM, but somehow VRAM still spills into RAM from the main opened application. They really need to fix this buggy AF VRAM management in the Linux NVIDIA driver.
Shared VRAM happens, but it’s just not as performant or as seamless as it is in Windows.
I think the main problem is that the driver and/or DXVK and friends cannot migrate memory between SysMem and VRAM, while Windows can do that. I think it was mentioned in DXVK issue reports once that this isn’t possible in DXVK without support from the kernel or driver. And thus, desktop apps crash unter VRAM pressure because the system has no way to migrate cold VRAM allocations away.
With such a system in place, it should be possible to limit the amount of VRAM usage for desktop to 512 MB - or rather guarantee it a maximum of 512 MB for somewhat decent performance while the rest would be dedicated to games then.
People often think that the driver cannot use SysMem because allocations fail if VRAM is full. But reality is that applications may explicitly ask for VRAM for performance reasons, and exactly that is failing. The driver would need to migrate memory away that is allowed to live in both regions. And this is what either cannot be done, or isn’t properly done by the driver.
Looking at the Xorg options description, I think what the driver CAN do is dual-allocating resources, and then discard the VRAM region on demand while keeping a shadow copy in SysMem. Then, if for performance reasons, it can copy the shadow back to VRAM, it will do that. Not sure if Wayland supports a similar thing. To me, it looks like it doesn’t: VRAM problems under Wayland are much more prominent and impactful… One of the reason why I still use Xorg (the other is global hotkeys because one important software doesn’t see hotkey presses under Wayland).
This issue persists with the 565 drivers.
Using nvidia-open drivers on Arch would sometimes crash with the same error on the desktop. Using nvidia instead of nvidia-open only causes it to happen with Arma 3 after playing for a couple of minutes, and Minecraft with shaders once in a while.
Here’s hoping this issue gets fixed.
UP
This made my 2GB Nvidia unusable in games on Linux - I moved to 8GB VRAM now but there many games that “leak vram” or go little over.
As everyone said - Windows does not haev this ussue.