Meta Platforms Inc. has introduced a new system called KernelEvolve, designed to significantly enhance the efficiency of artificial intelligence (AI) models across various hardware platforms. This innovation aims to address the increasing complexity associated with deploying AI models, particularly in the context of optimizing performance for different types of hardware, including NVIDIA and AMD GPUs as well as Meta's own MTIA silicon chips.
The KernelEvolve system employs a search-based methodology for kernel optimization, enabling it to autonomously create and refine multiple kernel candidates much quicker than traditional methods, which often required extensive engineering input. This automation has resulted in notable performance gains, achieving over a 60% increase in inference throughput for the Andromeda Ads model on NVIDIA GPUs and more than a 25% improvement for ads models on MTIA silicon.
Furthermore, KernelEvolve generates kernels in high-level domain-specific languages and translates them into lower-level programming languages, ensuring compatibility with various hardware architectures. It integrates an advanced knowledge base that provides relevant documentation, which is essential for optimizing code for proprietary chips like the MTIA. This dynamic adaptability allows KernelEvolve to remain effective amid evolving hardware technologies.