Lisa Su Reaffirms Commitment To Improving AMD ROCm Support, Engaging The Community

[AKSF_Ackermann]

he is the famous PS3 hacker afterall

Yea, nope.
Almost all of his work on that is taking fail0verflow's work and releasing it.

[timofonic]

Seeing is believing. Too much promises, I'll wait for reality.

[Spajdrex]

Hi Lisa, we love you!

[ermo]

Great, I have some faith In Lisa, so I hope the compute thing will get better for Linux.

Personally I would love to buy a AMD GPU, but the blender HW Raytracing acceleration isn't there yet and Nvidia is really good at it. But hope that the software side will get some serious attention now.

Just want to point out that I successfully installed the entire rocm stack on Solus yesterday, which allowed the upstream blender-3.5.1 release tarball's Cycles renderer to work with my relatively old RX Vega 64 in HIP GPU Compute mode. The classroom benchmark finished in 1m24s on the Vega 64 vs. 7m40s on ye olde R7 2700X.

Just because it's tedious to get working on some distributions, doesn't mean it's intractable (credit to the Solus staff who tested and merged the requisite stack work for this to become a reality). A new Solus release ISO is apparently coming soon, so even never hardware should work ootb.

[qarium]

Lisa Su should join phoronix.com we talked about this multible times in the past

and the answer was always in 2017 AMD had no money to pay the development for any compute product outside of the HPC supercomputer CDNA world.

i bought 6 vega64 in 2017 for compute workloads and i tried to install HIP/ROCm multible times and tried it with Blender and result is always:

Blender error message:
"No compatible GPU found for cycles
requires a AMD GPU with Vega or RDNA Archiecture
AMD driver version 22.10 or newer"

I do have Mesa23 installed on my Fedora 38

I do have the ROCm packaged installed with DNF

also a clear message should be helpfull is my vega64 to old to be supported do i need a new card ?

or will it be supported in the near future ?

in 2017 AMD had no money to pay for the compute development and now in 2023 they do have the money ? looks like Lisa Su want to say us this.

why can't we not have MESA compute like Glover ? instead of ROCm Glover in MESA just works.

if ROCm is not installed by default in the linux distros it is useless it need to be installed by default like MESA Glover.

[qarium]

Just want to point out that I successfully installed the entire rocm stack on Solus yesterday, which allowed the upstream blender-3.5.1 release tarball's Cycles renderer to work with my relatively old RX Vega 64 in HIP GPU Compute mode. The classroom benchmark finished in 1m24s on the Vega 64 vs. 7m40s on ye olde R7 2700X.
Just because it's tedious to get working on some distributions, doesn't mean it's intractable (credit to the Solus staff who tested and merged the requisite stack work for this to become a reality). A new Solus release ISO is apparently coming soon, so even never hardware should work ootb.

$ sudo dnf install rocm*

does not work on Fedora38... blender cycles does not work.

[X_m7]

In my case the ROCm OpenCL and HIP runtimes are on the Arch repos as rocm-opencl-runtime and hip-runtime-amd, at least for my RX 6650M the OpenCL side works fine with Folding@Home after working around a libc incompatibility with the F@H GPU core, and the HIP side works with Blender after I select the right Cycles render device in Blender's preferences (the Classroom render finished in 1m30s).

It doesn't seem to work with the Radeon 680M iGPU unfortunately, even clinfo returns an error somewhere in the middle for the 680M side, while F@H and Blender just crash if I try making them use the 680M, in fact suspend breaks after I try those things, while I guess it's not especially relevant for my case it seems like wasted potential for sure.

[bug77]

This is good news!

For who?
You and I can't install reaffirmed commitments. It's a mostly usage cry: "rats, please don't leave this boat, we swear we'll plug the holes in it".

In the words of master Yoda: "Do! Or do not!"
"Reaffirming commitment" is nothing in this context.

[bug77]

In my case the ROCm OpenCL and HIP runtimes are on the Arch repos as rocm-opencl-runtime and hip-runtime-amd, at least for my RX 6650M the OpenCL side works fine with Folding@Home after working around a libc incompatibility with the F@H GPU core, and the HIP side works with Blender after I select the right Cycles render device in Blender's preferences (the Classroom render finished in 1m30s).

It doesn't seem to work with the Radeon 680M iGPU unfortunately, even clinfo returns an error somewhere in the middle for the 680M side, while F@H and Blender just crash if I try making them use the 680M, in fact suspend breaks after I try those things, while I guess it's not especially relevant for my case it seems like wasted potential for sure.

And that's the best case scenario you're describing, using a distro with up-to-date everything. Step away from rolling releases for a moment and try to see which of the others can run ROCm easily.

[abott]

isn't about morality but about resources.

They make roughly $2 Billion in profit a quarter. That's billion, aka 1000 Million.

Plus, you think a software and chip company would have enough software and chips to make sure their shit works. GPU's are finally basically there, thanks to some people reading this. But how can your whole other stack, including their damn website, suck so much?

It's just incompetence, honestly. Especially from their software side. Meanwhile, the company with minimal software issues was just announce to have a trillion dollar evaluation. All because their software portion doesn't suck.