AI agent beats NVIDIA's FlashMLA by 9.8% on B200 after 22-hour black-box optimization run
Equivalent-Grass-527 · reddit · 2026-11-02
OpenBMB's ForgeKernel Live ran 22 hours of autonomous GPU kernel optimization on an NVIDIA B200 — with the agent working against a black box: it could execute and benchmark the official FlashMLA but had no access to source, docs, or profiling data.
Key numbers:
- 22 hours autonomous, 32 kernel versions across 42 iteration rounds (3 failed)
- First beat the baseline in under 10 hours (version 13)
- Final result 1.098×: 472µs vs 518µs, a 9.8% speedup (chart shows the first gen-1 kernel started at just 0.971×, climbing from 0.3× in the first hours)
- Zero human-written lines; 22 GPU-hours on one B200
They've open-sourced the optimized kernel plus 27 operators covering attention, Gated DeltaNet, Kimi Delta Attention, GEMM, and convolution, and released the ForgeKernel harness so others can run their own kernel-optimization experiments. The shift: from AI writing code to AI iteratively optimizing low-level GPU code against real performance measurements.
More from Infra
- T3 Code claims 2x OpenAI's token throughput — mocked for conflating routing with inference — scaling01 · 2026-11-02
- CI tip: GitHub Actions rounds each job up to a full minute — merge short jobs to save cost — DanielLockyer · 2026-11-02
- UBS: Musk's Terafab rethinks fabs — wafers spend only ~25% of cycle time processing — BenBajarin · 2026-11-02
- Unsloth's 1-bit DeepSeek-V4.1-Flash GGUF beats other quants by 10%+ on agentic benchmarks — solyarisoftware · 2026-11-02
- Musk: superintelligence compute will 'round up to 100%' be in space — elonmusk · 2026-11-02
- Qualcomm demos first-gen HBC silicon: 18x LPDDR5X bandwidth, 5x HBM4 per watt — ryanshrout · 2026-11-02