Strix Halo 8060S llama.cpp: Qwen3-30B-A3B UD-Q4_K_XL pp512 755.14 tok/s (amdvlk, depth 130560)
755.14 tok/s per-request prefill (llama-bench pp512, batch 1, depth 130560)
Fields
id
benchmarks/shhl-8060s-qwen3-30b-a3b-udq4-amdvlk-pp512-d130560Metric
prefill_tok_s
Value
755.14
Unit
tok/s per-request prefill (llama-bench pp512, batch 1, depth 130560)
Model
Qwen3-30B-A3B UD-Q4_K_XL GGUF
Methodology
THIRD-PARTY MEASURED (strixhalo-homelab.d7.wtf AI/llamacpp-performance, accessed 2026-10-16). llama.cpp llama-bench on a Framework Desktop (Ryzen AI Max+ 395 / Radeon 8060S, gfx1151), model unsloth Qwen3-30B-A3B-UD-Q4_K_XL.gguf. Vulkan AMDVLK 2025.Q2.1. llama-bench -fa 1 -r 1 --mmap 0 -d 130560. Single run (-r 1). Per-request batch-1, NOT aggregate.
Measured by
third_party
Reproducible
yes
notes
UD-Q4_K_XL is Unsloth dynamic Q4_K with an importance matrix; mapped to gguf-q4-k-m because there is no separate UD quant id. These AMDVLK and tuned-ROCm columns sit next to the already-filed plain-ROCm, hipBLASLt, and RADV columns on the same page. AMDVLK holds decode at long context (81.79 tok/s at depth 0, 85.11 at depth 130560) where plain ROCm collapses.
Notes
UD-Q4_K_XL is Unsloth dynamic Q4_K with an importance matrix; mapped to gguf-q4-k-m because there is no separate UD quant id. These AMDVLK and tuned-ROCm columns sit next to the already-filed plain-ROCm, hipBLASLt, and RADV columns on the same page. AMDVLK holds decode at long context (81.79 tok/s at depth 0, 85.11 at depth 130560) where plain ROCm collapses.