Meta Introduces KernelEvolve to Optimize AI Infrastructure via Autonomous Agents
Meta has unveiled KernelEvolve, an agentic kernel authoring system designed to optimize the low-level infrastructure supporting its AI models. As part of the Ranking Engineer Agent series, this tool addresses the scalability challenges of tuning software kernels for Meta’s heterogeneous hardware fleet, which includes NVIDIA and AMD GPUs, CPUs, and custom MTIA silicon chips. By treating kernel optimization as a search problem, KernelEvolve leverages Large Language Models to automate the profiling, optimization, and debugging processes that previously required weeks of expert engineering time, compressing them into hours. The system demonstrates significant performance gains, achieving over 60% inference throughput improvement for the Andromeda Ads model on NVIDIA GPUs and more than 25% training throughput improvement on MTIA chips. KernelEvolve supports various high-level domain-specific languages and low-level coding standards, ensuring broad applicability across different hardware architectures. This innovation aims to eliminate bottlenecks in AI infrastructure development, allowing engineers to focus on higher-level tasks while maintaining efficient model execution at scale. The underlying research is scheduled for presentation at the 53rd International Symposium on Computer Architecture in 2026.
Wire timeline
Meta Introduces KernelEvolve to Optimize AI Infrastructure via Autonomous Agents
Meta has unveiled KernelEvolve, an agentic kernel authoring system designed to optimize the low-level infrastructure supporting its AI models. As part of the Ranking Engineer Agent series, this tool addresses the scalability challenges of tuning software kernels for Meta’s heterogeneous hardware fleet, which includes NVIDIA and AMD GPUs, CPUs, and custom MTIA silicon chips. By treating kernel optimization as a search problem, KernelEvolve leverages Large Language Models to automate the profiling, optimization, and debugging processes that previously required weeks of expert engineering time, compressing them into hours. The system demonstrates significant performance gains, achieving over 60% inference throughput improvement for the Andromeda Ads model on NVIDIA GPUs and more than 25% training throughput improvement on MTIA chips. KernelEvolve supports various high-level domain-specific languages and low-level coding standards, ensuring broad applicability across different hardware architectures. This innovation aims to eliminate bottlenecks in AI infrastructure development, allowing engineers to focus on higher-level tasks while maintaining efficient model execution at scale. The underlying research is scheduled for presentation at the 53rd International Symposium on Computer Architecture in 2026.
Engineering at Meta