Kernel agents lose 62 points off Nvidia

AutoKernel’s success rate fell from 87% on Nvidia to 25% on another accelerator, a 62-point portability loss, in the 210-operator KernelGenBench study; specialized agents also averaged 5.11 million tokens per successful kernel. The benchmark consumed more than 15 billion tokens across six hardware platforms, making generated-kernel cost and portability measurable rather than incidental. Nvidia’s chip-design agent loop still depends on verified outputs, and operators should qualify generated kernels—and their generation bill—on every accelerator they intend to run.