Hi, I’m about to put together a 7950X system when the final part I’m waiting on arrives (should have been yesterday). It will have a Radeon RX 6800 GPU.
I’m a recent convert to Fedora from Debian. I had difficulties with getting an Nvidia 3050 working (on two systems - one laptop, one desktop) on Fedora 37 but achieved some success (not the laptop, alas) with help from this forum.
I was advised that Fedora Project is the definitive source for information on getting graphics drivers e.g. working with Fedora.
What is the accepted method of getting the correct AMD/Radeon drivers working with Fedora, specifically 37? I finally got my Nvidia installation working with rpmfusion. Is there a similar process for AMD? Or do I just use the amdgpu package that normally comes with Linux? As I say, I’m a recent convert to Fedora so don’t know if there are different solutions for getting AMD dGPUs operational. For the Nvidia drivers I followed a couple of articles I found on the web; one of them ultimately worked, but I ended up doing several installations from scratch until I succeeded.
I’m not anticipating the same level of frustration - and many re-installs - that I experienced with Nvidia GPUs, but I’m just trying to do some due diligence before I put the system together and start installing software.
Thanks. Presumably you’ ve proven the 6900XT is operating correctly. I have another system - 5950X with RX 6600 running Debian 11 - which has the standard amdgpu driver. glmark2 tells me that it’s using the 6600 yet Blender tells me that it can’t find a card that supports CUDA/OpenCL. I don’t know why this is (the 6600 is sufficient for Blender I believe), whether my system is not correctly configured. I’ve read that on Debian particularly, I might need to download AMD GPU firmware, but also that the firmware that’s available for the stable repo doesn’t support my card (this is one of the issues with Debian LTS - they’re well-known to be significantly behind available hardware). Output in dmesg leads me to believe the 6600 is being configured correctly, but I still might be missing the correct firmware. I realize that Fedora might not have the same issues as Debian in this regard, which is one reason I’m putting Fedora on new systems.
Oh, ok - I’m using it for gaming; which works perfectly. For Blender/productivity, you’ll need to install some extra packages. I think rocm-runtime and rocm-opencl?
As I understand it CUDA is for nvidia GPUs (at least according to wikipedia)
I do not use blender nor do I have any AMD GPUs, but is it possible blender will only work on nvidia? I would not assume that but without trying cannot answer.
Fedora is likely ahead of debian on firmware since they seem to update the linux-firmware package monthly if there is reason to do so.
It is. AMD GPUs employ OpenCL. The point was that Blender isn’t finding any GPUs capable of parallel compute tasks via any of the industry-standard methods. The 5950X system only has that one card and no iGPU obviously. I’m at a loss to explain why Blender can’t find a capable GPU given that the driver seems to be installed and working correctly. Not having fully capable firmware is the only thing I’m aware of that might explain it.
I believe you’re correct on this too, but have yet to prove it to my own satisfaction. That will only happen when I put the new machine together. I’m trying to ensure I have the best outcome once I put the OS on it and hopefully don’t have to go scrambling for drivers or firmware (and using backports like on Debian, which I’ve not done for GPU firmware, only kernel bits. Generally I’m OK with experimenting with, or even building kernels and drivers, but I get squeamish around firmware and BIOS updates - don’t want to brick anything!).
Fedora tries to keep up to date with firmware and the user normally has nothing to do in that regard. During boot the device is identified and the appropriate driver and firmware are loaded.
This normally only fails when the user has a very new device that has not yet received driver/firmware support.
In /lib/firmware/amdgpu I see a dozen different firmware files for the sienna-chichlid chipset that RX 6800 card uses. My linux-firmware package is version 20221109-144.fc37
Edit:
I just did an update on one of my VMs and with a dnf update it installed upgraded firmware to amd-gpu-firmware-20221214-145.fc37 which surely should support that newer AMD GPU.
That’s certainly newer firmware (if the 20221214 part indicates a date) than that which is available in the Debian 11 LTS release, and which pre-dates my card (it’s also not on the list of cards officially supported by the firmware). Hopefully within the next few days the system will be assembled, although bad weather has seemingly stopped parcel deliveries…
Blender 3.0+ for AMD support requires ROCm HIP (OpenCL was enough before that) Blender - ArchWiki
Fedora packages are too new to be supported by AMD proprietary drivers (those are intended for some LTS distributions).
For open-source solutions there’s an AMD employee working (in his spare time) on adding rocm-hip to Fedora, you can contact him and help with testing: Revisiting ROCm packaging - devel - Fedora Mailing-Lists
Read that topic (there’s some copr repo at the end, reported to be enabling blender to work on AMD GPUs):
Thanks for the pointer. Sounds like I might be forced to go to Ubuntu after all! Fedora too new, Debian too old. Ubuntu = Goldilocks? (I’m not really an Ubuntu fan however.)
Seemingly AlmaLinux is a possibility: it’s a continuation of CentOS and is deployed widely at CERN. I may take a look at that if I can’t make Fedora do what I want. Thanks for the link.