Strix Halo

DanielBowman

DanielBowman

What is the status of putting FreeBSD on Strix Halo (AMD Ryzen AI Max+ 395, usually with 128GB unified RAM) for LLM inference?
Clankers keep dissuading me from using FreeBSD. Ubuntu and Fedora, maybe Cachy are recommended systems.

It should work as per This table in forums from May

How well, tho? Have anyone has experience?
2uz36q-1968930515.jpg
 
No NPU.
No RocM.
on FreeBSD whatsoever.

AMD works closely with Ubuntu, FreeBSD ignored.
DRM trailing behind Linux optimizations (costing 5 to 15%)

Only pain, no gain. Zfs and jails are irrelevant on this setup as one should not run anything else but llama.cpp there for max memory.
 
DanielBowman rocm is not available natively but installs fine on compat/linux (plus with a more recent mesa version 26.2 vs 25.2), I'm not sure if everything necessary is exposed for OpenCL to work correctly by default, but vulkan does work (and yes I know there is a performance deficit).

What exactly did you test?
 
fahrenheit said:
What exactly did you test?
Click to expand...
I'm afraid I don't have the know how/skills/time to make proper testing i.e. my conclusions would be largely irrelevant.

I have no need for FreeBSD specialties (jail, zfs) on the Strix box, thus given the obvious hurdles, I would opt for Ubuntu 26.
 
You must log in or register to reply here.
Back
Top