kyuz0 Strix Halo 8060S NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL ROCm 7 nightlies 32k pp2048
353.69 tok/s
Fields
id
benchmarks/kyuz0-8060s-nvidia-nemotron-3-super-120b-a12b-ud-q4-k-xl-rocm7n-32k-pp2048Metric
prefill_tok_s
Value
353.69
Unit
tok/s
Model
NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL
Methodology
kyuz0 amd-strix-halo-toolboxes llama-bench sweep (results.json generated 2026-05-18, Fedora 43 kernel 6.19.12-200, llama.cpp build 1a68ec937 / b9193). ROCm 7 nightlies, -fa 1 -ngl 99 --mmap 0. Long-context cell: context depth 32768 tokens, prompt processing of 2048 tokens. ONE request, batch size 1, so the number is per-request, not aggregate across concurrent requests. Mean of the repeated runs; std 0.07 tok/s. Model file 78.02 GiB, 120.67 B params, quant Q4_K_XL.
Measured by
third_party
Reproducible
yes
notes
Not in the earlier kyuz0 filing, which covered only the default-context pp512/tg128 cells. Same results.json, the longctx32768 and longctx65536 rows.
Notes
Not in the earlier kyuz0 filing, which covered only the default-context pp512/tg128 cells. Same results.json, the longctx32768 and longctx65536 rows.