Meta's AI infrastructure team has a new tool for the part of deep learning work nobody talks about: writing the low-level kernel code that actually runs on the hardware.
KernelEvolve is an agentic framework that takes a kernel specification as input and produces optimized kernel code for heterogeneous hardware — NVIDIA GPUs, AMD GPUs, and Meta's in-house AI accelerators. It operates across multiple programming layers, from Triton and CuTe DSL down to hardware-agnostic low-level languages. The system uses a graph-based search process with a fitness function and retrieval-augmented prompt synthesis to adapt its output to the specific hardware and runtime context. Meta validated the system on the KernelBench suite, where it hit a 100% pass rate across all 250 problems at three difficulty levels and across 160 PyTorch ATen operators on three hardware platforms.
The significance here is less about the benchmark numbers and more about what this automates. Writing hand-tuned kernels for recommendation model training is slow, specialized work — the kind that can take engineers weeks per model variant. Doing that across a fleet of heterogeneous hardware accelerators, including custom silicon that no third-party tool supports, compounds the problem. KernelEvolve collapses that timeline to hours and lowers the barrier to bringing new AI chips into production.
Meta is not the first to chase automated kernel generation — Google's work on kernel auto-tuning and projects like OpenAI's Triton have pushed in similar directions — but deploying this at production scale across custom in-house silicon is the part worth watching. If the claims hold in production, this is less a research curiosity and more an infrastructure bet that custom hardware stops being a bottleneck only engineers can fix.