CAKE Paper: Compiler-Agent Co-Design Achieves 2.05x Speedup on Blackwell Kernels

hsu_byron · x · 2026-08-14

The CAKE paper introduces a compiler-agent co-design paradigm where the compiler is part of an evolving harness. Its IR is distilled by agents from production kernels and co-evolves with them. On B200, CAKE achieves 1.144x FlashML performance at 80M token budget vs 0.928x for direct CUDA/PTX. Agent-generated Kimi Delta Attention kernels achieve 2.05x geomean speedup over official FlashKDA.

Original post →

More from Infra

Infra channel →