Backend Performance Asymmetries on AMD Strix Halo for LLM Inference: A Case Study of ROCm/HIP vs. Mesa RADV Vulkan on gfx1151

The AMD Strix Halo platform (Ryzen AI MAX+ 395 with a Radeon 8060S iGPU on the gfx1151 RDNA 3.5 ISA and 128 GiB of unified memory) offers an unusual combination for solo LLM developers: enough VRAM to load 27-35B parameter models without quantization, at approximately one third of the cost of an equivalent NVIDIA-class workstation. The hardware ships with two viable GPU inference backends in llama.cpp: AMD's official ROCm/HIP stack, and Mesa's open-source RADV Vulkan driver with cooperative-matrix support. We benchmark both at an identical source commit on the Qwen3.6-35B-A3B model and find a precision-dependent asymmetry: Vulkan outperforms ROCm by 19-22 percent on Q4_K_M decode, while ROCm outperforms Vulkan by 117-121 percent on BF16 decode. The asymmetry is attributable to a single missing capability (bf16:0) in the Mesa STRIX_HALO Vulkan driver. Separately, we show that GGML_HIP_ROCWMMA_FATTN=OFF — a flag that AMD's published RDNA 3.5 best-practices recommends enabling — in fact yields up to 145 percent higher prefill throughput at 8K context on gfx1151. We release the full benchmark dataset and reproduction recipes under MIT.