CAKE
Compiler-agent co-design for evolving frontier GPU kernels.
GPU kernel agents typically see the compiler as a black box: they submit code and receive errors, correctness results, and timings. That interface hides the hardware scheduling structure an agent needs to diagnose and improve a kernel.
CAKE co-designs the agent and compiler around CAKE IR, a typed, hardware-explicit schedule representation. It exposes warp roles, memory movement, synchronization, and pipelines while giving agents verification, cost modeling, and localized diagnostics. The surrounding harness evolves as well, turning recurring failures into verifier rules, IR primitives, model calibrations, and reusable optimization tactics.
Reported results include:
- a best Flash-KMeans candidate reaching 1.144× the tuned FlashML baseline on B200
- agent-generated Kimi Delta Attention achieving a 2.05× geometric-mean speedup over official FlashKDA
- dispatcher-backed KNN and KMeans improving performance by 1.42×–2.12× across more than 400 shapes
CAKE targets NVIDIA GPUs from Ampere through Blackwell and separates single-shape kernel evolution from library-level generalization and dispatch.