Strix Halo 8060S llama.cpp HIP b8156: gpt-oss-120b MXFP4 token generation at depth 131072
25.26 tok/s (tg depth 131072, per-request batch-1)
Fields
id
benchmarks/strixwiki-8060s-gptoss-120b-mxfp4-hip-d131072-tgMetric
decode_tok_s
Value
25.26
Unit
tok/s (tg depth 131072, per-request batch-1)
Model
gpt-oss-120b MXFP4
Methodology
THIRD-PARTY MEASURED (Strix Halo HomeLab wiki llama.cpp performance page, reproducing the lhl/strix-halo-testing sweep). Framework Desktop, 128GB unified memory, ROCm/HIP backend, llama.cpp build b8156, -ngl 999, mmap disabled. gpt-oss-120b MXFP4. Depth 131072 is the prompt length the point was measured at (pp and tg swept together, not the default pp512/tg128 pair). Per-request batch-1, NOT aggregate. Wiki rounds the raw logs.
Measured by
third_party
Reproducible
yes