Strix Halo 8060S Vulkan/AMDVLK: gpt-oss-120b decode_tok_s (prompt 2048, gen 32, depth 0)
50.05 tok/s tg32 at context depth 0, batch 1, per-request
Fields
id
benchmarks/lhl-strix-8060s-gptoss120b-amdvlk-tg32-d0Metric
decode_tok_s
Value
50.05
Unit
tok/s tg32 at context depth 0, batch 1, per-request
Model
gpt-oss-120b
Methodology
lhl llama-bench b6608 (commit 1d108eef7), Vulkan backend, AMDVLK 2025.Q3.3 (2.0.355), GMKtec EVO-X2, Ryzen AI MAX+ 395, 128 GB RAM with 96 GB carved out as VRAM, Linux 6.16.4, performance profile. Model unsloth gpt-oss-120b-F16.gguf (65.4 GB; MXFP4 MoE weights, Q8_0 attention). -fa 1 --no-mmap -ngl 999. prompt 2048, gen 32, depth 0. Single request, NOT aggregate. Same machine family as the short-context lhl rows; this file is the 2025-10-03 depth sweep.
Measured by
third_party
Reproducible
yes
notes
Depth-0 pp2048 729.10 falls to 137.08 at depth 32k (~5.3x); tg32 falls 50.05 to 31.54. Per-request. The iGPU and the SoC are the same device; both ids are attached so either lookup hits.
Notes
Depth-0 pp2048 729.10 falls to 137.08 at depth 32k (~5.3x); tg32 falls 50.05 to 31.54. Per-request. The iGPU and the SoC are the same device; both ids are attached so either lookup hits.