Skip to content
← All posts

KDA generalizes to hardware it was never tuned for

ASM, HIP and ROCm — the loop finds optimization opportunities on less popular, less documented hardware too, which is the property that matters for the parts that do not exist yet.

Everything published about KDA so far has been on NVIDIA hardware with excellent documentation, a mature profiler and a large public corpus of prior kernels. That is a reasonable place to start and a bad place to stop, because it leaves open the obvious objection: perhaps the agent is retrieving, not reasoning.

So we pointed it at hardware where there is much less to retrieve. On ASM, HIP and ROCm — including AMD parts with substantially thinner documentation and far fewer published kernels — KDA still finds the optimization opportunity and still lands the speedup. Two of them are in FlyDSL, as pull requests #711 and #685.

This is the property that decides whether any of this is durable. New accelerators arrive with no corpus at all: NVIDIA's Rubin parts and Groq's hardware will both, at some point, be something an agent has read nothing about. A loop that works only where the answers already exist on the internet is a search engine with extra steps. A loop that works on the thin documentation is a method.

KDA · mit-han-lab/kernel-design-agents