AMD Improving The Linux Experience When Running New GPUs Without Proper Driver Support

[phoronix]

Phoronix: AMD Improving The Linux Experience When Running New GPUs Without Proper Driver Support

While AMD provided upstream open-source driver support for the Radeon RX 7900 series launch, the initial user experience can be less than desirable if running a new Radeon GPU but initially running an out-of-date kernel or lacking the necessary firmware support. With a new patch series posted AMD is looking to improve the experience by being able to more easily fallback to the firmware frame-buffer when their AMDGPU kernel graphics driver fails to properly load...

https://www.phoronix.com/news/AMD-Re-FW-FB-On-Failed-Probe

[schmidtbag]

Sounds like RDNA2 is already silently legacy.

Kinda, but RDNA2 at this point is old enough that most modern LTS distros probably support it by now. As I see it, the benefit of this is so people with cutting-edge hardware but stable/LTS kernels can still have a usable system so they can figure out how to get their GPU fully functional without having to transplant a different GPU or resort to an iGPU.

[rene]

Finally. I wondered for a decade why failing graphic drivers need to ruin otherwise working simple system frame buffer graphic, ..! :-/

[middy]

Kinda, but RDNA2 at this point is old enough that most modern LTS distros probably support it by now. As I see it, the benefit of this is so people with cutting-edge hardware but stable/LTS kernels can still have a usable system so they can figure out how to get their GPU fully functional without having to transplant a different GPU or resort to an iGPU.

pretty much if the distro has 5.11 (technically 5.10) or higher rdna2 will work. with 5.15 being a LTS its pretty much a guarantee. the only new thing that be missing is mesh shader support that came with kernel 6.1 (gang support) but you need mesa 23 for it to work and that won't be released till next year.

[V1tol]

Finally. I wondered for a decade why failing graphic drivers need to ruin otherwise working simple system frame buffer graphic, ..! :-/

And the fix was just in 30 lines of code lol.

[abott]

Thankfully this is much less of an issue back when I was running a Day 1 R9 270X and had daily issues for 3 years, including 9 months of hard locking 1-3 times a day lol. Good times, good times.

Now if something doesn't work it's usually just the small stuff, state tracking, hardware difference/unimplemented/changed stuff that just needs worked through. Hopefully it'll be worked out ahead and new issues can get rolled through 4-6 months before. With Intel floating their own new driver rework, maybe RadeonSI should get a facelift and rework to get rid of the cruft. NGG, different chips for compute, and SRIOV existing in only select places (but should be everywhere soon) I could see a rework being beneficial to keeping everything coherent, extend features so they can work sooner rather than later. It's not like, an issue, it's working well now. But eventually they'll have to de-platform RadeonSI and we should probably worry about that sooner rather than later. (SR-IOV)

Although I will say this, Vulkan is pretty nice and the other RADV stack is baller. Vulkan has shown why they should do it open. And it's not even as open as it could be, hopefully they'll get that out closer to release, too.

[ernstp]

"However the problem is further exaggerated in the case of amdgpu because
it has migrated to "IP discovery" where amdgpu will attempt to load
on "ALL" AMD GPUs even if the driver is missing support for IP blocks
contained in that GPU.
...
The perfect example of this is Ubuntu 21.10 and the new dGPUs just
launched by AMD. The installation media ships with kernel 5.19 (which
has IP discovery) but the amdgpu support for those IP blocks landed in
kernel 6.0. The matching linux-firmware was released after 21.10's launch.
The screen will freeze without nomodeset. Even if a user manages to install
and then upgrades to kernel 6.0 after install they'll still have the
problem of missing firmware, and the same experience."

So the new IP discovery made this situation a bit worse... even though you could hit this before also in a couple of ways.

[Mushoz]

Michael, did you have any luck running tensorflow-rocm with those 7900 cards? I am unable to do so, but I am unsure whether it's simply my technically unsupported rocm platform (Archlinux), or if Rocm simply doesn't work for tensorflow on these cards yet. Relevent Github issue: https://github.com/RadeonOpenCompute/ROCm/issues/1880