DuoDiscover AI Beats Human Engineers on NVIDIA SOL-ExecBench

A research group from Google Cloud, Stanford and Berkeley launched DuoDiscover, a tool that automates the discovery and optimization of technical engineering systems by integrating testing into its discovery process. DuoDiscover ranks #1 on NVIDIA's SOL-ExecBench and outperforms the best existing human and AI designed CPU prefetchers, achieving 2x the performance of experts on SWE-fficiency and gains of 2.68x over production TPU kernels. Additional results showed performance increases of 1.92x on vLLM and 1.44x on FreeToken.