Wanted to say that 570.133.07 was just released(at least on the Open drivers), and finally, System Monitor shows the correct total VRAM. Hoping this means that applications will be able to properly distribute memory instead of just crashing.
VRAM allocations also seem to stutter the whole system: Allocating GPU memory seems to be a system-wide blocking operation · Issue #164 · NVIDIA/egl-wayland · GitHub. This may explain why games become unresponsive when GPU is out of VRAM.
5600x 4070ti here, have terrible frames when vram gets to cap
nobara wayland plasma
endeavour wayland plasma
endeavour x11 plasma
pretty easy to reproduce, and the absolute worse offender is DCS World Standalone on a busy server. 3-12fps, 30 at best.. it should be over 160.
other games are Ready or Not, Star Citizen.
of note, setting dxgi.maxDeviceMemory = # to a value below your real cap(~2gb low), fixes the problem. (well, as much as having less vram is a ‘fix’, but the performance is playable again.) issue is: this doesnt always work. I have yet to be able to successfully force this onto DCS World Standalone, but it works great for star citizen however.
been like this for over 4 months, i think over 6. many games are unplayable at all, such as DCS, due to the only known workaround not working for it.
This issue is still present. If only Nvidia would create ram sharing in its drivers. Oh what a world would it be
It’s not RAM sharing that they need to implement, that already exists. It’s RAM eviction that they need to implement, so the driver would migrate idle/non-idle or performance-critical memory dynamically back and forth between device-local and host memory.
It’s frustrating how NVIDIA didn’t respond in 1.4 years…
Same issue here. Badly optimized games like Final Fantasy 7 Rebirth (that gobbles up every bit of vram it can), everything runs fine until it reaches VRAM limit and starts stuttering.
RTX 4070 Super
Noted on:
CachyOS, Wayland, KDE plasma
Nobara, Wayland, KDE plasma
Unfortunately, limiting vram with configuration file doesn’t work either.
Same here. I’m using an RTX 4060 (8GB) and noticing a massive VRAM overhead on Linux compared to Windows, specially on Directx12. In Silent Hill 2 Remake, using DX11 (DXVK) keeps allocation around 7.5GB with smooth performance. However, switching to DX12 (VKD3D) causes VRAM to hit the 8GB limit almost instantly. Even on the Low preset, performance tanks due to VRAM swapping and stuttering.
The same issue occurs in Final Fantasy 7 Rebirth and Icarus (Icarus, no matter if it is DX12 or DX11). S.T.A.L.K.E.R. 2 runs perfectly on Max settings, using around 7-7.5GB VRAM.
OS: CachyOS, 6.19.9-2 | Driver: 595.58.03
Seems to have been “improved” or rather “resolved” on some systems. On a laptop with an RTX 3060 6GB STALKER 2 would always go into single digit numbers whenever VRAM gets full, with driver 595.58.03 I am unable to reproduce this behavior, same applies to a fully maxed out Diablo 2 Resurrected. Will continue test some more and report back.
That was the reason I created the GreenBoost kernel modules: to orchestrate memory between T1 GPU VRAM and T2 system RAM while keeping all tensor computations on the tensor cores (with no CPU spillover). These modules work alongside normal NVIDIA kernel modules.
Today I’ve just released v2.8 after deleting NVidia name from the project (since this is not an official NVidia project), due law I needed to do the change
I’m accidented and had not in mind release the new version yet
v2.8 includes a new vulkan layer and GreenBoost Proton;
