Skip to content

Conversation

@pmenzel
Copy link
Collaborator

@pmenzel pmenzel commented Aug 5, 2026

Nvidia’s proprietary driver 580.178.04 supports CUDA 13, and our server GPUs.

Host(s) GPU Arch
aros (×2), heulsuse GeForce GTX 980 Ti, GM200 Maxwell
gameboycolor, hactar (×2) Tesla V100 PCIe 16/32GB, GV100GL Volta
beebee-ate (×4) GeForce RTX 2080 Ti, TU102 Turing
sigusr2 GeForce RTX 2080, TU104 Turing
bandersnatch, jabberwocky, theredqueen, whiterabbit A100 PCIe 40GB, GA100 Ampere
privateidaho A10, GA102GL Ampere

donald and others added 5 commits February 28, 2026 13:42
580.xx is the last branch supporting Maxwell, Pascal and Volta. From 590.xx
on those GPUs are listed under the legacy driver instead, and the fleet still
has GeForce GTX 980 Ti (GM200, Maxwell) in aros and heulsuse and Tesla V100
(GV100, Volta) in gameboycolor and hactar. So 580.xx is the newest branch
that can be deployed for all cards.

The GPU coverage is otherwise the same as in 570.144 — both start at Maxwell,
Kepler stays on 470.xx — and the supported CUDA version grows from 12.8 to
13.0.

Install libnvidia-nvvm70.so.4 explicitly. It is new in 580 and, like
libnvidia-api.so.1, carries no 'PKGVERSION', so the loop over
lib*.so.$PKGVERSION does not pick it up.

https://download.nvidia.com/XFree86/Linux-x86_64/580.178.04/README/
570.144, used for the 6.12.x kernels, cannot be used here. Linux 6.18 dropped
the deprecated EXTRA_CFLAGS from kbuild, and 570.144 still sets its include
paths with it, so the conftests pass but the compilation then fails because
the driver's own headers are no longer found. 580.178.04 uses ccflags-y and
builds all five modules cleanly.
Sign in to join this conversation on GitHub.

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants