kyuz0 Strix Halo 8060S gpt-oss-20b-mxfp4 ROCm 7.2.3 pp512
1786.17 tok/s
Fields
id
benchmarks/kyuz0-8060s-gpt-oss-20b-mxfp4-rocm-723-pp512Metric
prefill_tok_s
Value
1,786.17
Unit
tok/s
Model
gpt-oss-20b-mxfp4
Methodology
kyuz0 amd-strix-halo-toolboxes llama-bench sweep (results.json 2026-05-16, Fedora 43 kernel 6.19.12, llama.cpp b9187/9193). ROCm 7.2.3, -fa 1 -ngl 99 --mmap 0, default ctx. pp512 batch-1 per-request.
Measured by
third_party
Reproducible
yes