Shared GPU memory (as described in that article) is just how Windows decided to solve the problem of oversubscription of VRAM. Linux solves it differently (looks like it just allocates what it needs in demand and uses GART to address it, but I would like to know more).
So I’m curious what you mean when you say you miss it. Are you having programs crash OOM when running on Linux? Because that shouldn’t be happening.
It’s not ideal to be relying on shared gpu mem anyway (at least in a dgpu scenario). Kinda like saying you have a preference on which crutches to use.
There is one game called Cities: Skylines 2 that always fills up my VRAM, so yeah, I’m getting an OOM, but on the VRAM (I have GTX 1660super with 6 gigs of VRAM and I have 32 gigs of system RAM). I encourage you to try playing this game with a moderately sized city and with this GPU.
Assuming C:S2 uses DX and you’re running it through proton/dxvk, it’s ultimately the Vulkan driver’s job to page to system memory correctly. This honestly sounds like you’re seeing a bug. In that circumstance, it shouldn’t crash, it should just hurt performance from all the paging.
I see a couple of older issues where people were seeing exactly this kind of issue with DXVK+Nvidia.
This old Witcher 3 one where they blamed it on Nvidia’s memory allocator not playing well with linux THP (transparent huge pages). Disabling THP was a workaround.
This other issue for several titles that were hitting memory alloc failures despite having tons of system memory, just as you describe. They try several workarounds, but ultimately they believe it was fixed by a driver update.
One other thing to try is, idk if you’re running the game in dx11 or dx12 mode, but apparently both exist. If it’s currently running in dx11 mode, try the launch flag -force-d3d12. If you’re already using dx12, maybe try swapping back to dx11. Good luck!
Well, it probably is. I’ll soon be getting a GPU with a bigger VRAM and putting that GPU into my home server for Jellyfin as a replacement for QuickSync (NVENC is better, imo).
Shared GPU memory (as described in that article) is just how Windows decided to solve the problem of oversubscription of VRAM. Linux solves it differently (looks like it just allocates what it needs in demand and uses GART to address it, but I would like to know more).
So I’m curious what you mean when you say you miss it. Are you having programs crash OOM when running on Linux? Because that shouldn’t be happening.
It’s not ideal to be relying on shared gpu mem anyway (at least in a dgpu scenario). Kinda like saying you have a preference on which crutches to use.
There is one game called Cities: Skylines 2 that always fills up my VRAM, so yeah, I’m getting an OOM, but on the VRAM (I have GTX 1660super with 6 gigs of VRAM and I have 32 gigs of system RAM). I encourage you to try playing this game with a moderately sized city and with this GPU.
Assuming C:S2 uses DX and you’re running it through proton/dxvk, it’s ultimately the Vulkan driver’s job to page to system memory correctly. This honestly sounds like you’re seeing a bug. In that circumstance, it shouldn’t crash, it should just hurt performance from all the paging. I see a couple of older issues where people were seeing exactly this kind of issue with DXVK+Nvidia.
One other thing to try is, idk if you’re running the game in dx11 or dx12 mode, but apparently both exist. If it’s currently running in dx11 mode, try the launch flag
-force-d3d12. If you’re already using dx12, maybe try swapping back to dx11. Good luck!Well, it probably is. I’ll soon be getting a GPU with a bigger VRAM and putting that GPU into my home server for Jellyfin as a replacement for QuickSync (NVENC is better, imo).