AMD Strix Halo (Ryzen AI Max+ 395) 128GB, llama.cpp Vulkan: gpt-oss-120b MXFP4 MoE decode tok/s
33.7 decode tok/s (single-stream; pp512 386.8)
Fields
id
benchmarks/comm-strix-halo-gptoss120b-mxfp4-decodeMetric
decode_tok_s
Value
33.7
Unit
decode tok/s (single-stream; pp512 386.8)
Model
gpt-oss-120b MXFP4 (MoE, 117B params / 5.1B active)
Methodology
THIRD-PARTY MEASURED (lhl/strix-halo-testing llm-bench, Vulkan AMDVLK). gpt-oss-120b MXFP4 MoE (117B total / ~5.1B active): 33.7 tok/s decode, pp512 386.8. ~64 GB footprint - easy fit in 128GB. Prior-run note of '46-50 tok/s gpt-oss-120b' is achievable on newer tuned llama.cpp/ROCm builds - treat 34-50 as the range depending on backend + flags. MoE with ~5B active params is the matched workload for ~256 GB/s unified memory: only the active experts' bytes are read per token.
Measured by
third_party
Reproducible
yes
notes
The headline Strix Halo number and the reason to buy the platform: gpt-oss-120b at ~34 tok/s is genuinely interactive on a $2k, 12.5W-idle box. This is the MoE sweet spot - for DENSE 70B the same hardware does only ~5 tok/s. Also measured: gpt-oss-20b MXFP4 47.1 t/s; Qwen3-32B Q8_0 dense 6.4 t/s.
Notes
The headline Strix Halo number and the reason to buy the platform: gpt-oss-120b at ~34 tok/s is genuinely interactive on a $2k, 12.5W-idle box. This is the MoE sweet spot - for DENSE 70B the same hardware does only ~5 tok/s. Also measured: gpt-oss-20b MXFP4 47.1 t/s; Qwen3-32B Q8_0 dense 6.4 t/s.