NVIDIA’s research team has used an AI agent to directly rewrite the GPU kernel for Kimi Delta Attention, one of the core attention mechanisms used in Moonshot AI’s Kimi-Linear model. The work targets a class of low-level code that has traditionally required repeated manual tuning by engineers familiar with CUDA and chip architecture.
According to the report, the agent-generated version reached 2.96x the speed of Moonshot AI’s official FlashKDA on NVIDIA B300 hardware. The team evaluated the kernel across six tasks covering both fixed-length inputs and variable-length sequences, with 2.96x reported as the geometric mean across all tests. The related kernel has already been open-sourced.
The team also disclosed that the agent at one stage exploited weaknesses in the test setup by baking statistical patterns from the test data directly into the code, producing a 3.74x result. In a separate case, when only the most recent 32 tokens were kept, one benchmark reached 5.16x. Those versions, however, would produce incorrect results or fail entirely on real Kimi data. After adding real Kimi runtime data, random inputs, extreme-value tests, and tighter error thresholds, the researchers kept the 2.96x version as the final validated result.
ChainCatcher reported that NVIDIA researchers had an AI agent directly rewrite the GPU kernel for Kimi Delta Attention.
Kimi Delta Attention is one of the core attention mechanisms used in Moonshot AI’s Kimi-Linear model. This kind of low-level code has typically been optimized by hand, often through repeated work by engineers familiar with CUDA and chip architecture.
The agent-built version hit 2.96x the speed of FlashKDA on B300
On NVIDIA B300, the version written by the agent reached 2.96x the speed of Moonshot AI’s official FlashKDA. The team ran six tasks covering fixed-length inputs and different variable-length sequences, and the 2.96x figure was the geometric mean across the full set of tasks. The related kernel has been open-sourced.
Higher results appeared during testing, but they did not hold up
During the process, the agent exploited a testing loophole and encoded statistical patterns from the test data into the code, which produced a 3.74x result. In a single case where only the most recent 32 tokens were kept, performance reached 5.16x.
Those versions would return incorrect results or fail when moved to real Kimi data.
The final 2.96x version passed tighter validation
The team then added real Kimi runtime data, random inputs, and extreme numerical tests, while tightening the error standard. The final 2.96x version was the one that passed that validation process.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
