lhl Strix Halo: gpt-oss-120b MXFP4 tg32 on Radeon 8060S (Vulkan AMDVLK 2025.Q2.1 (-ub 512))
50.05 tok/s (tg32, batch 1)
Fields
id
benchmarks/lhl-strix-8060s-gpt-oss-120b-vulkan-amdvlk-tg32-d0Metric
decode_tok_s
Value
50.05
Unit
tok/s (tg32, batch 1)
Model
gpt-oss-120b-mxfp4
Methodology
lhl/strix-halo-testing DGX Spark vs Strix Halo comparison (Framework Desktop, Arch 6.17, all amd_iommu/tuned opts). llama.cpp build 6792, gpt-oss-120b MXFP4, backend Vulkan AMDVLK 2025.Q2.1 (-ub 512), tg32 at context depth 0. Per-request batch-1. DGX Spark column from ggerganov's #16578 sweep (build 6767, CUDA).
Measured by
third_party
Reproducible
yes
notes
Tuned rocWMMA branch (lhl/llama.cpp rocm-wmma-tune) reaches 36.43 tg32 at d32768 vs 17.82 stock — the best known long-context Strix Halo config.
Notes
Tuned rocWMMA branch (lhl/llama.cpp rocm-wmma-tune) reaches 36.43 tg32 at d32768 vs 17.82 stock — the best known long-context Strix Halo config.