inference/research
verifiedmedium · 0.70updated 2026-10-14

Strix Halo 8060S Vulkan RADV gpt-oss-120b MXFP4: tg128 tok/s

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

Fields

id
benchmarks/ky0-gpt-oss-120b-mxfp4-vulkan-radv-tg128
Metric
decode_tok_s
Value
56.61
Unit
tok/s (tg128, per-request batch-1)
Model
gpt-oss-120b
Methodology
THIRD-PARTY MEASURED (kyuz0/amd-strix-halo-toolboxes results.json). Ryzen AI Max+ 395 + Radeon 8060S (gfx1151), 128GB UMA, Fedora 43 kernel 6.19.12. llama.cpp build 9187/9193, env 'vulkan_radv' (Vulkan RADV), -fa 1 -ngl 99 --mmap 0, pp512 + tg128 mean over repeats, per-request batch-1. Model gpt-oss-120b (117B, 0 GiB).
Measured by
third_party
Reproducible
yes

Cross-references

Accelerators (2)

Engine (1)

Interconnect (1)