AI Agent Beats NVIDIA FlashMLA Baseline by 9.8% on B200 GPU
An autonomous AI agent developed by OpenBMB outperformed NVIDIA's FlashMLA baseline by 9.8% on a B200 GPU during a 22-hour black-box optimization run. The agent successfully accelerated the kernel execution time from 518 microseconds down to 472 microseconds without having access to the source code. This achievement marks a significant technical milestone by demonstrating that AI agents can autonomously optimize low-level hardware code purely through performance feedback rather than direct code analysis. It highlights a shift toward AI-driven hardware acceleration and kernel optimization in complex computing systems. During the 22-hour run, the agent tested 32 kernel versions across 42 iteration rounds and beat the baseline for the first time at version 13. OpenBMB has open-sourced the optimized kernel alongside 27 operators and the ForgeKernel harness for community experimentation.
## BACKGROUND
GPU kernels are specialized functions executed on graphics processing units to accelerate heavy computations like deep learning tasks. FlashMLA is an efficient attention mechanism library optimized for high-performance model inference, while the NVIDIA B200 is an advanced data center GPU built on the Blackwell architecture designed for extreme AI workloads.