MLPerf v6.1 Atlas_Inference Strix Halo (8060S): Qwen3.6-27B NVFP4 SingleStream median TPOT
47.28 ms median TPOT (agentic SingleStream)
Fields
id
benchmarks/mlperf-v6-1-atlas-strix-halo-qwen36-27b-tpotMetric
itl_ms
Value
47.28
Unit
ms median TPOT (agentic SingleStream)
Model
Qwen3.6-27B-NVFP4
Methodology
MLPerf v6.1 closed Atlas_Inference agentic submission: 1007 turns BFCL, single-stream, temp 0, wall 7108.6s. Strix Halo Ryzen AI Max+ (Radeon 8060S gfx1151), ROCm core-7.13, atlas engine stack @eabfa8f. TPOT median 47.28ms (~21.1 tok/s), TTFT median 2713ms, TPS 10.22. Per-request.
Measured by
third_party
Reproducible
yes
notes
Agentic single-stream: TPOT med 47.3ms ≈ 21.1 decode tok/s on a 27B dense NVFP4. BFCL accuracy 86.23.
Notes
Agentic single-stream: TPOT med 47.3ms ≈ 21.1 decode tok/s on a 27B dense NVFP4. BFCL accuracy 86.23.