Strix Halo 8060S Vulkan RADV Qwen3.6-35B-A3B UD-Q4_K_XL depth-0 decode

Protocol deterministic-prime-extend-completions-v1, runner local-llm-speed v1, measured at the API endpoint of an OpenAI-completions server. Mean of 2 repeats after 1 warmup. Depth 0: no primed prefix, so the whole prompt is computed. Follow-up prompt 2048 tokens, 128 generated tokens, KV cache type default, speculative decoding off. Prefill is effective follow-up prompt processing: uncached tokens divided by client-observed time to first output. Decode is client-observed generation throughput over the tokens after the first, so TTFT is excluded. Run run-d58dcca093daefad, completed 2026-09-02. Hardware: AMD Ryzen AI Max+ 395, Radeon 8060S iGPU (Strix Halo). llama.cpp, backend Vulkan RADV, container image llama-vulkan-radv. Weights: unsloth/Qwen3.6-35B-A3B-MTP-GGUF file Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf. Per-request, concurrency 1, not aggregate.

Concurrency 1 is one request, so the figure is per-request. The same run also measured concurrency 4, 8 and 16; those are aggregate and are not recorded here. Not comparable with the repo llama-bench pp512/tg128 rows: different prompt length and a server endpoint rather than llama-bench.