TomAoki1d ago

Latest Production Branch of #NVIDIA #GPU #driver package 595.58.03 is landed on main (aka latest) branch of #FreeBSD ports tree.
https://cgit.freebsd.org/ports/commit/?id=b35026e786a2c761d4b2b7853a989ab23dda5e9b

No plan for merging into quarterly branch (now 2026Q1), but next quarterly 2026Q2 would be created early April.

Note that this version drops support for a bunch of old (pre-Turing generation of architectures) GPUs.
Anyone using affected GPUs need switching to newly created -580 variant of legacy branch.

You can find your GPU is still supported or not here.
https://us.download.nvidia.com/XFree86/FreeBSD-x86_64/595.58.03/README/supportedchips.html

580 series are NOT yet found on NVIDIA Unix Driver Archive page as Legacy version,
https://www.nvidia.com/en-us/drivers/unix/

but I think one would appear later, as 580.xx is listed as driver for legacy GPUs.

ports - FreeBSD ports tree

Show threadLaurent Cimon1d ago
@TomAoki I've got an RTX 5070 running those drivers, but they feel a bit overkill. Since there is no CUDA or AAA games, what do people use the latest Nvidia drivers for on FreeBSD?
Show threadTomAoki

@clf
Maybe not work well for every use-cases of CUDA, but CUDA can be used via Linuxulator on FreeBSD.
Install ports / pkgs below in addition with drivers and Linuxulator (now defaulted to rl9) and start the programs wanted via `nv-sglrun` would allow CUDA to (partially) work. Note that anything really needs nvidia-uvm.ko (not available on FreeBSD version) shouldn't run sanely.

emulators/libc6-shim
devel/libepoll-shim
science/linux-ai-ml-env
x11/linux-nvidia-libs

Example:
% nv-sglrun nvidia-smi
/usr/local/lib/libc6-shim/libc6.so: shim init
Sat Mar 28 12:15:50 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.58.03 Driver Version: 595.58.03 CUDA Version: 13.2 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA RTX A400 Off | 00000000:01:00.0 On | N/A |
| 30% 59C P5 N/A / 50W | 846MiB / 4094MiB | 26% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| No running processes found |
+-----------------------------------------------------------------------------------------+

Mar 28 at 3:17amWeb
Show threadLaurent Cimon1d ago
@TomAoki interesting. So I could try doing some ML work without booting Linux? You say that it doesn't work 100%, would PyTorch be stable?
Show threadTomAoki1d ago

@clf
Not sure.
What I've confirming on upgrades is to run nvidia-smi with and without nv-sglrun and see CUDA is detected with nv-sglrun used. And run clpeak (benchmarks/clpeak) to confirm if CUDA works via nv-sglrun or not.
Not actually using anything others like PyTorch myself.

As far as I've heared, FreeBSD version lacks nvidia-uvm.ko, which should be required for CUDA to fully functional using UVM (Unified Virtual Memory), because of the lack / incompatibilities of implementations in FreeBSD.

science/linux-ai-ml-env has dummy-uvm.so that (IIUC) mimics as if nvidia-uvm.ko is loaded, but is actually a dummy. Anything confirming the existence of the feature but not actually using the feature would work, but anything really require the functionalities wouldn't.

Show threadLaurent Cimon1d ago
@TomAoki I wish I weren't too busy to try it, this semester's taking a toll on me and I need to direct my energy towards my courses. But this summer I'm probably going to try this and I'll post here and hashtag FreeBSD with the results.