ForgeKernel tops all CANNBench hardware tracks with anonymous submissions

Entries submitted under the name ForgeKernel now sit at the top of all three hardware tracks of the public CANNBench leaderboard, which evaluates AI-generated Ascend operators. The leaderboard scores submissions on compilation, functional correctness and performance, weighting them 0.2, 0.3 and 0.5 respectively. Pandaily checked the official-tasks package v2.0.0 on October 27, 2026 and recorded ForgeKernel’s 950PR score of 3,649.86 (49 of 60 operators passing, 980 of 1,200 cases cleared, 1.876× speedup) Pandaily AI & Tech China.

The same submitter leads the 910C track with 4,338.67 points (57 operators passing, 1.894× speedup) and the 950DT track with 3,899.93 points. All three entries were updated on the same day and carry no hidden-evaluation score, which applies only to the older 1.1 task package.

The leaderboard does not disclose the developer; three different usernames filed the entries under the organisation label ForgeKernel, with no associated company or project page. Chinese outlets such as Leiphone, IT Home and Yicai reproduced the story, and IT Home noted a claim that the kernels were written end-to-end by AI without line-by-line human tuning. Pandaily could not verify that claim nor the reported Nvidia-focused benchmark results.

Our read: The anonymous success on CANNBench underscores how open-source benchmarking can surface high-performance kernels without revealing their creators, and suggests that AI-assisted code generation may already be competitive on Ascend hardware, though the lack of independent verification limits confidence in the broader claims.