ForgeKernel kernel generator outperforms experts on NVIDIA Hopper and Blackwell

Researchers introduce ForgeKernel, a kernel-generation system designed to accumulate and reuse knowledge across optimization runs while enabling the continued evolution of kernel architectures. The system analyzes hardware behavior from reference kernels, distills validated architectural decisions, and uses them as references for subsequent optimization iterations. According to the evaluation, ForgeKernel matches or outperforms expert references on NVIDIA Hopper and Blackwell GPUs, and ranks first on several CANN-Bench solution leaderboards. Open article →