ForgeKernel reportedly beats DeepSeek FlashMLA by 9.8%
QuantumBit reported that ModelBest’s unattended ForgeKernel agent generated a FlashMLA Prefill kernel for an NVIDIA B200 without consulting DeepSeek’s implementation. The livestreamed run used 42 iterations over 22.5 hours, with the final kernel taking 472 microseconds versus 518 microseconds for FlashMLA, a reported 9.8% improvement. QuantumBit also said ForgeKernel subsequently became open source through the Intelligent Evolution Laboratory and the OpenBMB community.
The result suggests an agent-based approach to kernel optimization may complement manually written, hardware-specific performance code, but its broader significance still depends on independent validation of the reported benchmark.
AI零帧起手写算子，反超DeepSeek人类手写巅峰：快了9.8%！ – 量子位
< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400">
AI零帧起手写算子，反超DeepSeek人类手写巅峰：快了9.8%！
之前还连登昇腾三榜第一
#ForgeKernel#ModelBest#DeepSeek#FlashMLA#NVIDIA B200#OpenBMB#CANNBench#Ascend