Meta KernelEvolve: Agentic Kernel Optimization Treats Code Generation as Search — 60% Throughput Gain in Hours

Engineering at Meta·high signal

Meta's KernelEvolve treats kernel optimization as a search problem rather than one-shot code generation, exploring hundreds of alternative Triton kernel implementations to find solutions matching or exceeding human expert performance. It achieved 60% ads model inference throughput improvement in hours (vs. weeks for humans), operating continuously across NVIDIA GPUs, AMD GPUs, and Meta's custom MTIA silicon serving trillions of daily inference requests. Paper accepted at ISCA 2026.