I bought a Strix Halo so you don't have to. Really, you don't
So, you wanna run big models, do ya? But you're on a budget, and Nvidia cards are ridiculous; Apple has taken all the big-ass M4 Max Studios off their shop; and a second-hand one with 512GB is selling for close to €20,000!
The solution, I thought, would be to cheat the system, go outside of the usual vendors, and pick up a GMKTec Evo X2. This bad boy features the impressive-sounding AMD AI Max+ 395 chip, also known, even more impressively, as Strix Halo. Even better, you can load it out with an incredible 126GB unified RAM. After applying a handing €700 coupon on Amazon, it cost about €3,000. Not nothing, but a damn sight cheaper than any of my other options.
The first thing I did with it was wipe out Windows and install Ubuntu. The second thing I did with it was install Ubuntu, again, because I realised there's a new LTS version, 26.04. The third thing I did was regretting all my life choices up until this point, as I hit multiple snags with the new Ubuntu, particularly boring-not-relevant-but-important-to-me issues with the new LUKS boot thingymajag.
I'm getting side-tracked, which is exactly what happens with every new Linux install on a desktop... you end up spending days working on shit that should just work, but mysteriously doesn't.
Add to that, we're running an AMD GPU, with shared memory, on Linux. I should have expected the pain.
Nvidia's CUDA is supported by all the AI frameworks, all the operating systems, and your cousin. AMD's Rocm? Not so much. Ubuntu does actually offer Rocm drivers, but they're already hopelessly out-of-date in its brand-new release.
Three days it took to get it running LLMs as fast as it can. Which is about 33% slower than my Mac, and only half as fast as my RTX 3090, both of which required absolutely no config. And after all that effort, the performance was, well, tepid.
I'm almost exclusively using Qwen3.5, 35B A3B on all my platforms, so it's pretty easy to compare.
Mac Studio M4 Max: About 75 tokens per second
RTX 3090: About 120 tokens per second
Strix Halo, unoptimised: About 10 tokens per second
Strix Halo, optimised: About 25 to 50 (depending on flavour of model)
In real-world conditions, there's a pretty noticeable jump from 50 to 120 tokens per second.
At first I was upset. Really upset. I thought I should return it, but I hate returning things to Amazon, knowing that they'll possibly end up in landfill. I spent a few days really mulling it over.
Finally I had an epiphany: this wasn't an AI machine. This was just a fast PC with a lot of RAM. It just happens to do okay at AI.
With that in mind, I've shifted my view and found a real use for this machine. It will become my primary dev box, freeing up my RTX 3090 for production stuff. And given prices of RAM at the moment, a fast PC with 128GB of RAM for €3,000 is actually a bloody good deal.
For the nerds, here are my notes on making Strix Halo run about as fast as it can.