lhl 8060S Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL Vulkan tg1 decode
15.16 tok/s
Fields
id
benchmarks/lhl-8060s-qwen3-235b-a22b-instruct-2507-ud-q3-k-xl-vulkan-tg1Metric
decode_tok_s
Value
15.16
Unit
tok/s
Model
Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL
Methodology
lhl strix-halo-testing llm-bench (github.com/lhl/strix-halo-testing, Qwen3-235B-A22B-Instruct-2507-UD-Q3_K_XL/results.jsonl). llama.cpp llama.cpp-vulkan, backend Vulkan, no -fa, ngl=99, tg1 batch-1 decode at depth 0 on Ryzen AI Max+ 395 (Radeon 8060S gfx1151). System RAM peak 104192MiB, avg power 17W, max temp 68.0C. Per-request batch-1.
Measured by
third_party
Reproducible
yes