speed-bench: Strix Halo ROCm results on a Minisforum MS-S1 Max
By alexdelprete · GitHub
- Source published
- Aug 25, 2026
- Test performed
- Aug 25, 2026
- Added to MiniPC.guru
- Sep 17, 2026
- Source reviewed
- Sep 17, 2026
- Tested hardware
- AMD Ryzen AI Max+ 395, reported by the contributor as a Minisforum MS-S1 Max · Radeon 8060S, gfx1151, with 124 GiB of GTT exposed through amd_iommu=off, amdgpu.gttsize=126976, ttm.pages_limit=32505856 and ttm.page_pool_size=32505856 · 128 GB LPDDR5X-8000, quad channel · not applicable: the model is fully resident in memory, with no SSD streaming, so token generation does not read from disk. The source does not state the SSD capacity, so this run is shown against both storage options
- Software / version / OS
- ds4 speed-bench, ROCm backend · ds4 built from main around commit 44e47de with make strix-halo; Ubuntu ROCm 7.1 packages, hipcc 7.1.1, plus the rocWMMA headers named in STRIXHALO.md · Ubuntu 26.04 LTS, kernel 7.0.0-30-generic
- Power profile
- firmware power mode Performance, about 130 W sustained package limit and a steady 131 W package draw during the run
- Test conditions
- DeepSeek-V4-Flash IQ2_XXS with w2Q2K, AProj Q8, SExp Q8 and Out Q8, chat v2 imatrix build of 31 July. Command exactly as published in speed-bench/README.md, reading promessi_sposi.txt, context swept from 2048 to 65536 in steps of 2048, 128 tokens generated, machine otherwise idle. This row is context 2048: prefill 153.17 tok/s, generation 16.10 tok/s
Editorial review checks the source and documentation; it does not certify the measurement or mean we tested the device ourselves.





