Fedora 41 Looks To Ship Upcoming AMD ROCm 6.2 For Latest AI Capabilities

[oleid]

It would interesting to see how big is the delta between the packages from the AMD repo and the ones from Debian repo.

It would seem that debian stable has version 5 of rocm. Hence, I need a newer version to support my card.
I just removed everything and now tried again cleanly. I realized my previous installation was not clean, as I installed the packages manually one-by-one instead of using a meta package.

Also, I used jammy=ubuntu-22.04 packages, not focal=ubuntu-20.04.
In essence, I used this package here to get the repo.

There is, however, a dependeny, which is not available. Namely libpython3.10, which is required by rocm-gdm.
I created a fake package using equivs. The next step was convincing APT to use the packages from the rocm repo, not the internal repo (since the package versions in the rocm repo are smaller ). Thus, I needed an override file:

/etc/apt/preferences.d/99rocm
-----
Package: rocminfo
Pin: release a=jammy
Pin-Priority: 990

Package: rocm-device-libs
Pin: release a=jammy
Pin-Priority: 990

Package: rocm-cmake
Pin: release a=jammy
Pin-Priority: 990

Package: hipcc
Pin: release a=jammy
Pin-Priority: 990

That appears to do the trick. After that I can simply run apt install rocm.

Also, did you install the dkms kernel module or just user space libraries?

Apparently, the dkms module is currently in use (although I didn't recall installing it).
Given the debian stable kernel is quite old, it might be required - I'm not sure.

And what GPU hardware did you use, if this is not a secret?

Not a secret. I got hands on a AMD Radeon RX 7900 XTX (due to its 24 GiB VRAM).

[KDE_FOREVER]

The sad thing is that in the end it will all be for naught.

As many of you know, I have positions in NVIDIA and am counting on them continuing to generate revenue and grow.

Having said that, I fully expect that eventually someone will come out with an discrete NPU that is many times faster than ROCm on AMD and CUDA on NVIDIA with significantly lower power consumption and effectively kill the AI on GPU craze that is fueling the AI stock surge that has been driving the tech sector gains.

Basically what happened with cryptomining only with AI.



If AMD was smart, a big IF in my opinion, they would be working on bringing such a device to data centers, because they are not going to out CUDA NVIDIA with ROCm.

you have positions only in local garbage cans where you collect bottles because your trolling for masters in redmond does not earn you enough for the rent

[contik32]

I used the ones from AMDs repo to get the most recent ones. TBH I didn't try the integrated ones. Afair I used on Debian stable.

It would interesting to see how big is the delta between the packages from the AMD repo and the ones from Debian repo.

Also, did you install the dkms kernel module or just user space libraries? And what GPU hardware did you use, if this is not a secret?

[oleid]

Is it ROCm packages from Debian repos or from AMD repos? Or built from sources?

I used the ones from AMDs repo to get the most recent ones. TBH I didn't try the integrated ones. Afair I used on Debian stable.

[contik32]

But you can use ROCm just fine on Debian stable (what I'm using currently). Speed is good as well. I'm looking forward to seeing what ROCm 6.2 is going to provide.

Is it ROCm packages from Debian repos or from AMD repos? Or built from sources?

[Mathias]

What does "Fed41 code freeze" mean? here is the roadmap for 41. Is it the "branch 41 from Rawhide" date (2024-08-13) or "Beta Freeze" (2024-09-17) or "Final Freeze" (2024-10-22)? I'm eagerly waiting for 6.2 for it's alleged Navi1 support... On the other hand, I'm not interested enough to compile from source - I wasted enough time trying to get it working. (FYI HSA_OVERRIDE_GFX_VERSION=10.3.0 doesn't work anymore)

[DiamondAngle]

They've half assed rocm for years, now they'll half ass ai on rocm allowing them to neglect everything else even more. Super exciting

Honestly ROCM on instinct hw is perfectly fine, it has all the features you would need, performance is good most of the time there is a rich set of libraries that all work well. There are some rough edges for sure, the profiling tools are a bit iffy sometimes and the occasional bug can creep into even the instinct parts and sometimes the performance dosent reach quite to the theoretical limits.

Besides cuda, rocm is the best high performance compute platform by far and that is an incredible increase in quality over any previous AMD effort which where... poor at best, its also light years ahead of oneAPI still, atho intel is slowly catching up too.

[sophisticles]

The sad thing is that in the end it will all be for naught.

As many of you know, I have positions in NVIDIA and am counting on them continuing to generate revenue and grow.

Having said that, I fully expect that eventually someone will come out with an discrete NPU that is many times faster than ROCm on AMD and CUDA on NVIDIA with significantly lower power consumption and effectively kill the AI on GPU craze that is fueling the AI stock surge that has been driving the tech sector gains.

Basically what happened with cryptomining only with AI.



If AMD was smart, a big IF in my opinion, they would be working on bringing such a device to data centers, because they are not going to out CUDA NVIDIA with ROCm.

[clockwork]

the irony of programmers destroying the own jobs

If you're replaceable by an LLM in any job, you're not good enough at your job. Especially as a programmer.

[oleid]

But you can use ROCm just fine on Debian stable (what I'm using currently). Speed is good as well. I'm looking forward to seeing what ROCm 6.2 is going to provide.