The Silicon Reckoning settled Vulkan over ROCm. Months later, a smaller question — does the RADV driver's own version matter? — returned +31.6% prefill on identical hardware, from a container, without touching the pinned enterprise host.
Ten models, two families, two drivers, one iGPU. The size sweep meant to map a smooth curve found a cliff instead — dense models fall off it, Mixture-of-Experts models do not. And the newer driver's gift turned out to have a price only dense models pay.
With Ollama disabled and GTT freed, the fleet reran the entire playoff under Vulkan RADV. qwen2.5:14b Q2 dropped from 113 seconds to 21 seconds. No language drift. gemma4:12b Q2 dropped from 93 seconds to 30 seconds. The hardware had been capable the whole time. The backend had not.
The fleet investigated why large models were timing out and why Q2 produced Thai text. The answer was not in the models. It was in the physics of the machine — an iGPU, unified memory, and two inference servers fighting over the same 34GB pool.