AMD's Ryzen AI Max+ 395 ("Strix Halo") puts a large Radeon 8060S GPU next to up to 128 GB of shared memory. On Windows, AMD says up to 96 GB of that can be set aside as graphics memory. That is more than any consumer graphics card — three times an RTX 5090.
What fits
Everything on this site. Models that do not reach 8-bit quality even on a 32 GB RTX 5090 — FLUX.2 dev, MiniMax H3 — fit here at 16-bit or 8-bit. Full list →
The catch: speed
Image and video sampling is mostly limited by raw compute, and the Radeon 8060S is a 40-compute-unit integrated GPU — far smaller than a big desktop card. Memory bandwidth is the second limit: LPDDR5X-8000 on a 256-bit bus gives about 256 GB/s by simple arithmetic, against 1792 GB/s on an RTX 5090 and 1008 GB/s on an RTX 4090. So each step takes much longer.
That makes it a machine for models that simply do not fit anywhere else, or for people who value "it runs" over "it runs fast". For FLUX.1-class models that fit on a 16 GB desktop GPU, the desktop GPU wins easily.
Things to check before buying
- The 96 GB figure needs the 128 GB version of the machine, and the graphics memory setting in AMD's software. (A newer Ryzen AI Max+ PRO 495 generation, shown in September 2026, supports up to 192 GB.)
- ComfyUI runs on AMD through ROCm; check that your model's custom nodes support it.
- FP8 files save memory but give no speed-up on AMD — with this much memory, just use Q8_0 or 16-bit.
Strix Halo vs RTX 5090, model by model →
Source
- AMD: "up to 96GB VGM available on a Ryzen AI MAX+ 395 128GB machine on Windows" — amd.com →