And you can have as much “VRAM” as you want by dynamically allocating it, just like your m2 mac mini does
@Adrian_Joachim
Can you show me?
I just swapped from the intel 11g to the AI 350 board, I have allocated 32G to the GPU (seems max the AMD software will do is %50 of mem or 48GB in my case, on windows)
LM Studio is showing it as VRAM.
For games you don’t need to do anything extra and for ai stuff idk bit it has been done before, even here on this forum.
I assume what they are talking about is GTT. It requires kernel 6.10+, and is memory the GPU can use but unlike “vram” it is not reserved for the GPU.
I have mine set to 64GB using kernel param amdgpu.gttsize=64000 (it will default to 50% of ram but I have 96GB of ram and wanted more than 48GB allowed to GPU) and I am able to load 50GB models entirely on the 780m of my 7840u.
Had to patch ollama to get it to support, so ymmv depending on what software you use.
2 Likes
Which distro are you using? And is this ROCm, or vulkan?
I run fedora 42 with kde, but as far as I know gtt should be available as long as your kernel version is 6.10 or over.