AI-Written Kernel ForgeKernel Tops NVIDIA and Ascend Operator Benchmarks

As of October 22, 2026, ForgeKernel, a set of operators entirely written by AI without manual line-by-line tuning, has topped both the NVIDIA SOL-ExecBench FlashInfer-Bench collection and the Ascend CANNBench A3 hardware leaderboard. The benchmark shows a baseline speedup of 1.835x, 25% ahead of the runner-up.

ForgeKernel also surpassed DeepSeek's open-source NVIDIA version of the DeepSeek v4.1 Flash operator, using an entirely AI-driven iterative optimization process. The result highlights the growing competitiveness of AI-driven operator optimization in public benchmarks.

Earlier, DeepSeek open-sourced its inference framework DeepInfer and released operator libraries including FlashMLA and DeepGEMM. A manually tuned version of FlashMLA achieved over 90% of hardware limits. ForgeKernel reset that benchmark within the same week, shifting operator efficiency competition from manual expertise to AI-driven iteration.

Operator efficiency directly affects how fast a model runs on a given chip and influences token costs. As competition in China's AI sector extends from model-level to compute-level, ForgeKernel's achievement offers a new technical reference.