Entries submitted under the name ForgeKernel now hold first place on every hardware track of CANNBench, the public leaderboard the CANN open-source community runs for AI-generated Ascend operators. CANNBench scores submissions by compiling them, checking numerical correctness and then timing each kernel on Ascend NPU runners, with weights of 0.2 for compilation, 0.3 for function and 0.5 for performance.
On the current official-tasks package, version 2.0.0, which covers 60 operators across five difficulty levels and 1,200 test cases, the leaderboard read as follows when Pandaily checked it on October 27. On the 950PR track, ForgeKernel scored 3,649.86 with 49 of the 60 operators passing, 980 of 1,200 cases cleared and a geometric mean speedup of 1.876 times the reference. Second place, an entry tagged CANNBot PyPTOPro S2 frontier, scored 3,552.34 while clearing more cases, 1,057, at a lower 1.181 times speedup.
ForgeKernel also leads the other two tracks: 4,338.67 on 910C, with 57 operators passing and a 1.894 times speedup, and 3,899.93 on 950DT. All three entries were updated on October 27 and carry no hidden-evaluation score, which the leaderboard applies to the older 1.1 task package.
What the leaderboard does not show is who built it. The three top entries are filed under the organisation label ForgeKernel by three different leaderboard usernames, with no company, laboratory or project page attached. Pandaily could not identify the developer from any public source.
Chinese tech outlets including Leiphone, IT Home and Yicai carried the same article about the project within days of each other, with IT Home marking its version as republished commercial information. That copy claims the kernels were written end to end by AI with no line-by-line human tuning, and that the same system also leads a separate Nvidia-focused benchmark. Those remain claims by the promoters rather than verified findings, and Pandaily is not repeating the comparison numbers attached to them.
What is checkable is narrower but more interesting: on an operator benchmark maintained by the Ascend software community, the current top scores on all three chip variants come from an anonymous submitter whose kernels trade case coverage for raw speed. CANNBench publishes per-operator logs for each run, so the result can be inspected, and overtaken, by anyone who signs up.
SEE ALSO: Huawei Spells Out CANNLab, CANNBot and CANNJudge as Its Open Ascend Developer Base · DeepSeek Open-Sources Ascend Versions of TileLang, DeepGEMM and DeepEP