I’m currently running Fedora 42 with CUDA 13.0, the proprietary NVIDIA drivers, and all kernel modules functioning correctly (including nvidia, nvidia_uvm, and nvidia_modeset). CUDA compilation tools (nvcc) and libraries (libdevice, nvvm) are working as expected.
Fedora 43 has just been released, but there is currently no official CUDA repository available at:
Before upgrading my system, I would like to know whether there is any estimated release date or planned timeline for when the cuda-fedora43.repo (or the Fedora 43 x86_64 directory) will become available.
Some additional context that may be relevant:
Fedora 43 ships with GNOME Wayland-only (X11 session removed), which may affect driver behavior, especially with NVIDIA’s EGLStreams and hybrid/Optimus configurations.
Kernel updates in Fedora typically require akmods to rebuild the proprietary driver; therefore, CUDA support depends on driver/kernel ABI compatibility.
I would like to avoid upgrading until the CUDA repository is officially published, to ensure that the NVIDIA driver stack, CUDA toolkit, and kernel modules remain stable.
Given all this, is there any information, roadmap, or internal scheduling regarding when Fedora 43 will be officially supported by CUDA, or when the cuda-fedora43.repo will be published?
Even an approximate timeframe (e.g., weeks vs. months) would be very helpful for planning.
Thank you for your attention and for the continuous support of the Linux ecosystem.
I waited several months before upgrading my desktop with NVIDIA drivers (for my RTX 4070 Ti Super) from Fedora 42 to 43, assuming it would be working by now. Since upgrading several days ago, my desktop graphics have been randomly crashing like five times a day. It seems to work fine for a while before the graphics crash in the middle of playing a video or something. Very sad.
Ironic that I bought an expensive NVIDIA GPU only for it to render my desktop virtually unusable.
My “solution” has been to not use my computer. Better than doing some work then having it crash within a few hours. :-(
I’m typing this on my Fedora 42 laptop, so I’m in the same boat twice. My upgraded desktop is unusable, while I’m deathly afraid to try to upgrade my laptop.
I don’t expect anyone to drop everything to fix this (esp. over the holidays), but it’s disappointing no one from NVIDIA has responded to this thread since it was opened November 13th.
For anyone like me whose Fedora build keeps crashing due to their NVIDIA GPU being unstable on Wayland, I just ran sudo dnf install @xfce-desktop-environment and logged into a (Wayland+X11) XFCE desktop instead of the default (Wayland-only) GNOME desktop on my Fedora 43 desktop. Has only been up 15 minutes, but video hasn’t yet crashed. #fingers_crossed Will hopefully buy me time till NVIDIA engineers get NVIDIA GPUs working better with Wayland.
To switch desktop, on the login screen, click your username but then immediately select the settings (gear) icon on the bottom right to switch your selected desktop to XFCE before typing in your password.
Hopefully NVIDIA will add official CUDA Toolkit support for Fedora 43 soon.
That said, this caught my eye, so wanted to clarify just in case it helps someone:
Are you a CUDA developer who needs CUDA Toolkit (as opposed to just the regular CUDA driver?)
If not, your best bet on Fedora is to not use the official NVIDIA repo. Instead, use the RPM Fusion packaged driver, as it is better integrated into Fedora: https://rpmfusion.org/Howto/NVIDIA
This is my daily workstation setup for some time now and the experience, even with Wayland, has been good in recent driver versions. This should work well for most people and there’s no need to go the X11/XFCE route (unless you just prefer XFCE or have a specific edge-case that works better on X11.)
If you ARE a CUDA developer (specifically require CUDA Toolkit), on the other hand, then IMO Fedora is not a great fit because on one hand, the NVIDIA driver works best while installed from RPM Fusion, and at the same time, you need the CUDA Toolkit from NVIDIA’s repo. There’s a configuration that can make this work OK most of the time but it is messy and may have occasional issues, e.g. if a new CUDA Toolkit version is released and requires a new NVIDIA driver version, but it is not available in RPM Fusion yet, etc. But if you still want to go this route then you should read the instructions and known issues here: https://rpmfusion.org/Howto/CUDA
I DO have nvidia-container-toolkit installed on my desktop from the nvidia-container-toolkit repo. I’m definitely NOT a CUDA developer, but I was using this to run local LLMs on this desktop.
When I set up this desktop in 2024, there were THREE different ways to set up NVIDIA GPUs (aargh!!!), and no one seemed to agree on the best solution. IIRC, I started with Nouveau first before quickly switching because NVIDIA seemed committed to making everything work nicely on Linux.
I’ve had things working pretty nicely, aside from having to rebuild some stuff after every upgrade, till my machine started crashing repeatedly after I upgraded to Fedora 43.
My XFCE desktop hasn’t crashed since I installed it yesterday (hooray!), so I’m happy enough now but will consider other approaches (starting with the one you mention here) if my needs change or my current setup starts crashing as my GNOME setup has been.
Apologies for hijacking this thread. Final update: My current theory is that my problem relates to my NVIDIA GPU wanting swap and my disabling of swap because this desktop also serves as a Kubernetes worker node. This desktop was pretty stable before upgrading from Fedora 42 to 43, but something about memory management could have changed. When I first built my K8s cluster, swapoff was required. It’s now configurable, so I’ll try enabling swap and hope that prevents these crashes. (The desktop has recently been crashing even when seemingly relatively idle, so swap may not be the problem, but it could be.)