[phdelightful]

I think I’d buy something with Strix Halo or Strix Point if there was official ROCm support. As of 6.4.1 from earlier this month there’s still not, as I understand it. I’d be delighted to be corrected on this matter.

[michaellarabel]

There is (unofficial) ROCm support for Strix Halo with ROCm 6.4.1. But like Llama.cpp and such were seg faulting but ROCR-based OpenCL was working and other workloads.

ROCm GPU Compute Performance With AMD Ryzen AI MAX+ "Strix Halo": https://www.phoronix.com/review/amd-strix-halo-rocm-benchmar...

[flakiness]

By the way thanks for working on this! I read all of your reviews on this device and it's been very informative.

[pantalaimon]

Does it work with RustiCL?

[perplexe]

What’s AMD's strategy in not having consumer chip support for ROCm? It's puzzling. No way to get critical mass of development interest if the bar to entry is high.

[_aavaa_]

Their plan appears to be TheRock [0]. Further open sourcing and engaging the community and leveraging that to expand support faster.

There are some recent discussions on YouTube about it [1], including one with a senior VP [2].

[0]: https://github.com/ROCm/TheRock [1]: https://www.youtube.com/watch?v=6tASUo7UqNw&t=4551 [2]: https://www.youtube.com/watch?v=0B8JOtS2Tew

[RandomRandy]

I think that part of the issue is the split between CDNA for data centers [1] and RDNA for consumer products [2] with AMD only having the money to focus on the bigger data center market. There are rumors that both architectures will be merged into UDNA in the future, which will hopefully improve ROCm support, but for now it's lacking

[1] https://www.amd.com/en/technologies/cdna.html [2] https://www.amd.com/en/technologies/rdna.html

[roenxi]

The strategy (seems to be) targeting data centres and focusing support efforts on the cards most likely to be used in one. There is an expectation that ROCm will work on pretty much everything but their drivers aren't good so in practice it is dicey whether it actually does.

[rcarmo]

You can run theoretically run ollama on it, as with the earlier APUs (I did it on an M780 by allocating 16 of the machine's 32GB to the iGPU). I am _very_ interested in getting my hands on one because I see it as a decent compromise between power, RAM capacity (with soldered-on RAM, it's got pretty good latency) and performance.