Hi to you all !
I’ve been using the nvidia and cuda drivers in their .run format, for some years now. While it still works fairly well for my use case on a Fedora 41 (xfce/kde) workstation, I was considering the rpm packages provided by the repo (this one I believe : Index of /compute/cuda/repos/fedora41).
Pros and cons ?
And how should I proceed for this transition please ? I would gladly follow any specific advices you might have.
Thank you for your help
for notes :
cuda .run driver is installed (I do not update that often) in /usr/local/cuda-xx with conf in /etc/ld.so.conf.d/ (+ ldconfig)
Pros: better OS management/integration.
Cons: the DC driver repos sometimes lag a version or two behind .run Unix driver (like at this very moment the latest .run Unix driver is 570.124.04 while the DC repos are stuck at 570.86.15).
To uninstall .run, just executing it with --uninstall should be sufficient.
To install from the DC repos, follow the official docs.
Hi @morgwai666,
Thank you for your answer. I assume you’re refering to both the cuda and nvidia driver. Then I just have to follow this after --uninstall I suppose :
For fedora : sudo dnf install nvidia-driver kmod-nvidia-latest-dkms cuda-toolkit ?
The thing I’m mostly concerned is the aftermath of uninstalling the drivers (leftover interfering conf files ?), but if there’s nothing to worry, I don’t mind being a little behind the .run and I’m using timeshift as a disaster recovery tool. I should be good to go.
It is possible to install both or just one of them: see further section in the doc.
that’s compute only without desktop components: not sure if that’s what you intended. Also unless you have some very old GPU, you should rather use “open” modules.
Honestly I’ve always been using the DC drivers, so no idea how clean uninstall of .run will be…
Now I’m getting confused, the doc (chapter 8.5.2. Desktop-only System driver-installation-guide) seems to refer to these packages as Desktop-only System ?
I need the cuda and nvidia drivers for Blender, mostly. I also use HW acceleration for podman (jellyfin)
If there are no difference performance-wise between open and proprietary kernel modules, I think I can use the open KM (as long as a RTX 2060 isn’t … too old ^^’ ?)
apologies: you mixed desktop-only packages with cuda-toolkit which confused me indeed. This mix will probably not work fully: at best it will automatically install all absolutely necessary compute dependencies resulting in a half-baked cuda support.
So assuming that you need cuda and desktop components (since you suggested desktop-only packages before), you should basically install the full driver by dnf install nvidia-open
not that I heard of.
it should be ok: open modules support GPUs back to Turing arch exactly (RTX-20xx and GTX-16xx on desktop).
Thank you morgwai666. That’s it, no more questions left ^^
That actually looks a little more straightforward to me now, and I will proceed in a few days.