“Other high-scoring entries on the KernelBench-Mega benchmark decompose the problem into 4 to 14 separate kernel launches per token.”