ForgeKernel reportedly surpasses DeepSeek kernels as AI optimization tops two benchmarks

What happened

LeiPhone reports that ForgeKernel, whose kernels were written entirely by AI without manual line-by-line tuning, topped two kernel benchmarks as of October 22. It ranked first in the SOL-ExecBench FlashInfer-Bench collection and the CANNBench A3 hardware leaderboard, reporting a 1.835× benchmark speedup and a 25% lead over second place on the latter. The report also says ForgeKernel surpassed DeepSeek’s open-source NVIDIA implementation for DeepSeek v4.1 Flash kernels during the week DeepSeek open-sourced DeepInfer.

Read next