ForgeKernel

ForgeKernel is an autonomous AI agent for writing and optimizing GPU kernels, shown publicly through ForgeKernel Live from OpenBMB and the AI4AI Lab on October 29, 2026. In the live demo, the agent builds a DeepSeek attention kernel from scratch and keeps improving it while viewers follow every iteration on a public dashboard.

Source: X trending story on ForgeKernel Live

Key points

1. What it does: the agent writes a kernel, compiles it, verifies correctness, benchmarks it, and uses the result to decide the next change, in a loop with no human edits.
2. Live and transparent: the dashboard shows graphs, logs and benchmarks for each iteration, compared against expert references such as FlashMLA.
3. Results so far: the trending summary reports the DeepSeek kernel's runtime moving from 593 µs to a figure given as "1.834" (units not stated) within hours, with Yuxuan Li of AI4AI Lab saying it neared top performance in about 4 hours.
4. Track record: ForgeKernel previously topped the NVIDIA SOL-ExecBench and Huawei CANN-Bench leaderboards, so it targets both NVIDIA and Ascend hardware.

Why it matters

Hand-tuned kernels are one of the scarcest skills in AI infrastructure. An agent that reliably reaches expert-level kernels could cut inference cost for new model architectures much faster.

How to follow or try it

• Watch ForgeKernel Live through OpenBMB's channels to see how the agent explores optimizations.
• Use it as a reference: compare its generated kernels with your own for the same operator, and check them with your test suite before deploying.
• Benchmarks to know: SOL-ExecBench (NVIDIA) and CANN-Bench (Huawei Ascend) are where its results are reported.

Always re-verify correctness and performance on your own hardware; agent-generated kernels can be tuned to a specific benchmark setup.