Qwen3.5 Benchmarks Split ROCm, Vulkan

Backend choice matters more than raw GPU class here. Qwen3.5 flips winners depending on whether you're in prefill, decode, or multi-GPU mode.

• –AMDVLK Vulkan beat ROCm on single-GPU generation for the 35B MoE and 27B dense models, with the biggest gains on the lightest active-compute path.
• –ROCm dominated 27B prompt processing by 3.5-4x, which makes it the better pick for long-context RAG, summarization, and document analysis.
• –The 122B dual-GPU run reverses the smaller-model pattern: ROCm wins both prefill and decode, so multi-GPU coordination finally pays for itself.
• –M5 Max looks genuinely strong for local inference because unified memory lets it dominate prefill without PCIe baggage.
• –W6800 is the cautionary tale: RDNA 2, ROCm incompatibility, and chipset x4 bandwidth can swamp any apples-to-apples GPU comparison.