inference/research
verifiedmedium · 0.75updated 2026-10-14

Strix Halo 8060S (Vulkan AMDVLK) gpt-oss 120B MXFP4: tg128 decode tok/s

33.7 tok/s (tg128, per-request batch-1)

Fields

id
benchmarks/strix-halo-8060s-gpt-oss-120b-mxfp4-amdvlk-tg128
Metric
decode_tok_s
Value
33.7
Unit
tok/s (tg128, per-request batch-1)
Model
gpt-oss 120B
Methodology
THIRD-PARTY MEASURED (lhl/strix-halo-testing llm-bench summary table). Hardware: Framework Desktop, AMD Ryzen AI Max+ 395 (Strix Halo) + Radeon 8060S iGPU, 128GB LPDDR5x-8000 unified memory. llama.cpp ~b5863, Vulkan AMDVLK backend, flags (default); llama-bench -fa 1 --mmap 0 -ngl 99, pp512 + tg128, per-request batch-1. Peak resident 64187 MiB. Numbers verbatim from the best-backend summary table.
Measured by
third_party
Reproducible
yes

Cross-references

Accelerators (2)

Engine (1)

Interconnect (1)