LLM speeds on AMD Strix Halo
Every language model measured on one machine: AMD Ryzen AI MAX+ 395, Radeon 8060S iGPU (gfx1151) and XDNA2 NPU, 128 GB unified memory, Nobara Linux 44. 48 models with a measured speed, 248 single measurements, and 5 models that were tried without producing a number.
Click a column header to sort. The numbers come from different harnesses and are not a ranking to the decimal — see how they were measured.
No row matches. That model was probably never tried on this machine.
How the numbers were measured
- A
- Smoke test through Ollama and Lemonade, 2026-09-15: six short German tasks, 400-token output cap, one run per model.
- B
- Candidate run, 2026-09-16: the same six tasks, 2000-token cap, thinking switched off.
- C
llama-bench,-ngl 999 -p 512 -n 128unless the configuration says otherwise, 2–3 repetitions.- D
llama-serveror the running service, fixed prompts, median or mean; with speculative decoding where stated.- E
- colibri, decode speed as reported by the engine.
Caveats
- A and B are single runs and good for orders of magnitude only.
- Speculative decoding (MTP, DSpark, EAGLE3) depends on the text; code and prose give different acceptance rates.
- Prompt speed depends on prompt length and micro-batch. Compare it only within one model.
- August HIP (ROCm) figures come from a build that computed wrong results on gfx1151. The speed was the same as that of the corrected build.
- The list says nothing about answer quality.