tested on 3060, debian 13. freeing 300 megs is easy, just disable gsp. 50 megs is thanks to an undocumented NVReg key, which may or may not work on different cards. (there are similar NVReg keys that didn't work well on my card, however could be useful on others: https://github.com/NVIDIA/open-gpu-kernel-modules/blob/main/... )
the other 50MiB is nvidia kernel module patching, 39 of which was a worst case memory reserve (32 general, and 7 for copy engine stuff). 9mb was reducing page table entry budget (12GB/4KB * 8bytes per PTE is 24megs, i couldnt go lower since it started breaking haha.). and there was around 4.5 megs of old legacy code.
i only tested llama.cpp since it's my only usecase.
locallmfan•1h ago
the other 50MiB is nvidia kernel module patching, 39 of which was a worst case memory reserve (32 general, and 7 for copy engine stuff). 9mb was reducing page table entry budget (12GB/4KB * 8bytes per PTE is 24megs, i couldnt go lower since it started breaking haha.). and there was around 4.5 megs of old legacy code.
i only tested llama.cpp since it's my only usecase.