# An AI agent just beat NVIDIA's FlashMLA baseline by 9.8% on B200, without seeing its source code — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/ee5f8c8929094e05

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-02 14:53 ingested 2026-11-03 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xlzaq0/an_ai_agent_just_beat_nvidias_flashmla_baseline/)

## retained text



```
OpenBMB just wrapped up ForgeKernel Live, a \~22-hour autonomous GPU kernel optimization run on NVIDIA B200.

The interesting part isn't just the speedup. The agent was optimizing against a black box. It could execute and benchmark the official FlashMLA implementation, but couldn't access its source code, documentation, or profiling information.

Some numbers from the run:

* 22 hours of autonomous optimization
* 32 kernel versions across 42 iteration rounds
* First beat the baseline in under 10 hours (version 13)
* Final result: **1.098× performance**
* 472 µs vs. 518 µs — a 9.8% speedup

They're now open-sourcing the optimized kernel, along with 27 operators covering attention, Gated DeltaNet, Kimi Delta Attention, GEMM, and convolution.

They've also released the ForgeKernel harness, so others can run their own kernel optimization experiments.

What's interesting is the shift from AI writing code to AI iteratively optimizing low-level GPU code against real performance measurements.

Curious how far this approach can go as these agents get better at CUDA and GPU architecture.

**Live run:** [https://openbmb.cn/forgekernel/](https://openbmb.cn/forgekernel/)  
**GitHub:** [https://github.com/OpenBMB/ForgeKernel](https://github.com/OpenBMB/ForgeKernel)
```



document fetched 2026-11-03 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xlzaq0/an_ai_agent_just_beat_nvidias_flashmla_baseline
