The competition among Chinese large language models is extending from the model layer to the compute layer: no matter how strong a model is, it ultimately runs on chips through operators. With the same model and the same chip, swapping in a better set of operators can sometimes deliver several-fold performance gains. Operator efficiency determines how fast the same GPU can run, and it also determines the price of each token.
DeepSeek recently fully open-sourced its inference framework DeepInfer. Following model weights and operator libraries such as FlashMLA and DeepGEMM, they have opened up the entire model execution pipeline. Among these, FlashMLA had a version hand-tuned line by line by engineers on the chip, with operator performance reaching over 90% of the hardware limit — the best result human craftsmanship has achieved so far on some chips.
Yet that limit was broken by AI in the same week. Reportedly, as of October 22, ForgeKernel — written entirely by AI with no manual line-by-line tuning — simultaneously topped two operator benchmarks: first place in NVIDIA's home turf SOL-ExecBench FlashInfer-Bench collection, and first place on the CANNBench A3 hardware leaderboard maintained in Ascend's official CANN repository, with a baseline speedup of 1.835×, leading second place by 25%. Beyond the rankings, AI's ability to autonomously optimize operators should not be underestimated: on the operators of DeepSeek v4.1 Flash, ForgeKernel also achieved a performance breakthrough over the NVIDIA version open-sourced by DeepSeek through AI iteration from 0.
This is a relay race in the evolution of intelligence. Whether it is DeepSeek's open-source spirit or the fierce competition on operator optimization leaderboards, everything ultimately points in the same direction: making Chinese AI run faster on every chip.
NVIDIA SOL-ExecBench FlashInfer-Bench
Ascend CANNBench A3 hardware leaderboard