Strix Halo homelab wiki — llama.cpp performance: Qwen3-30B-A3B UD-Q4_K_XL depth sweep (HIP / Vulkan RADV / Vulkan AMDVLK)

Third-party measured llama-bench depth sweep published on the Strix Halo HomeLab wiki llama.cpp performance page, run by lhl on a Framework Desktop (Ryzen AI Max+ 395 / Radeon 8060S, gfx1151). Model is unsloth Qwen3-30B-A3B-UD-Q4_K_XL.gguf. Three backends: HIP (TheRock ROCm 7.0.0-0, hipBLASLt), Vulkan RADV (Mesa 25.1.6), Vulkan AMDVLK (2025.Q2.1). Command shape: llama-bench -m <gguf> -fa 1 --mmap 0 -ngl 999 -d <depth>, five repeats, reported as average tok/s. HIP uses -ub 2048; RADV -ub 1024; AMDVLK -ub 512. The page also records the same shape on Qwen3-235B-A22B-UD-Q3_K_XL, which this source does not quote numbers from. Distinct from sources/strix-halo-homelab-wiki-llama-cpp-performance (the page as a setup guide) and from sources/lhl-strix-halo-testing-strix-vs-spark-readme (gpt-oss-120b vs DGX Spark). Page re-fetched 2026-10-16; tables reproduced verbatim.