llama.cpp GPT-OSS 120B MXFP4 MoE batch-1 decode on AMD Radeon 8060S (Strix Halo)
49.0 tok/s
Fields
id
benchmarks/llamacpp-strix-halo-8060s-gpt-oss-120b-mxfp4-decodeMetric
decode_tok_s
Value
49
Unit
tok/s
Model
gpt-oss-120b
Methodology
llama-cli-rocm (llama.cpp b10440, commit 6b4344ecc) on Ubuntu 26.04 LTS (kernel 7.0), native ROCm 7.x targeting gfx1151 on GMKtec EVO-X2 (AMD Ryzen AI Max+ 395, Radeon 8060S 40 CUs, 128GB LPDDR5X with ttm.pages_limit=30146560). Flags: --model gpt-oss-120b-MXFP4.gguf --n-gpu-layers 99 --ctx-size 8192 --flash-attn on --jinja --temp 1.0 --top-p 1.0 --top-k 0 -p <prompt> -n 512 / -n 2000. Measured steady-state batch-1 token generation at 49.0 tok/s (prompt processing 165.9 tok/s).
Measured by
third_party
Reproducible
yes
notes
Demonstrates high throughput on unified memory: 120B MoE achieves 49 tok/s because only active experts are evaluated per token, requiring ~4-5 GB of weights streamed per token rather than the entire 120B parameter set.
Notes
Demonstrates high throughput on unified memory: 120B MoE achieves 49 tok/s because only active experts are evaluated per token, requiring ~4-5 GB of weights streamed per token rather than the entire 120B parameter set.