ROCm

AMD's inference optimizations work well one at a time. The gap opens when they run together, and closing it needs test hardware more than it needs missing APIs.