CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution

Letting AI agents write GPU kernels in a checkable compiler language beats writing raw CUDA.

1.144x

tuned baseline

Best Flash-KMeans kernel speed

Same agent writing raw CUDA/PTX only reached 0.928x

How it works

Agents get thin feedback
Kernel agents see only errors, pass/fail and timing, so they can't tell which decision caused a hang or stall.
Typed IR plus verifier
Agents write Cake IR, an explicit schedule of warp roles, buffers and barriers. The compiler checks it and pinpoints problems.
Faster kernels, less time
Cake IR runs beat a tuned Triton baseline in half the evolve time. Agent-built Kimi Delta Attention runs 2.05x faster.

From the paper