ForgeKernel, an AI-written operator, ranks first on NVIDIA and Huawei kernel benchmarks
Article timeline
The articles in this story. Times are ET.
OpenBMBOfficialOpenBMB praises ForgeKernel, a self-evolving kernel optimization harnessAIOpenBMB's post follows up on JustGoal with a self-evolving kernel optimization harness, ForgeKernel, part of the Forge series. Per the quoted post by @_LiYuXuan_, ForgeKernel ranked #1 on every leaderboard tested, including NVIDIA's SOL-ExecBench (FlashInfer-Bench) and HUAWEI's CANN-Bench. A paper on ForgeKernel is forthcoming.
- PandailyNewsForgeKernel tops CANNBench's Ascend operator leaderboard on all three tracks
AIEntries under the name ForgeKernel hold first place on all three hardware tracks of CANNBench, the CANN community's public leaderboard for AI-generated Ascend operators, as of October 27. On the 950PR track, ForgeKernel scores 3,649.86 with 49 of 60 operators passing and a 1.876 times geometric mean speedup, ahead of CANNBot PyPTOPro S2 frontier at 3,552.34. The three top entries carry only the ForgeKernel organisation label, and Pandaily could not identify the developer.
- LeiphoneNewsForgeKernel, an AI-written operator, tops two kernel benchmarks after DeepSeek's DeepInfer release
AIDeepSeek recently open-sourced its DeepInfer inference framework, following its model weights and operator libraries such as FlashMLA and DeepGEMM. In the same week, ForgeKernel, an operator written entirely by AI with no manual line-by-line tuning, ranked first on NVIDIA's SOL-ExecBench FlashInfer-Bench collection and on the Ascend CANNBench A3 hardware leaderboard as of October 22, with a 1.835× benchmark speedup, 25% ahead of second place.