Race to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen... - Ajit Mathews

Race to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogeneous Hardware - Ajit Mathews, Meta AI workloads are diversifying faster than any single hardware platform can keep up. This session explores how Meta is advancing PyTorch as the unifying interface for heterogeneous hardware. We’ll show how Triton and Torch-Inductor enable fast, portable kernel development across GPUs and MTIA, and how AI-generated kernels are accelerating developer efficiency. At Meta, we’re building the software rails that turn hardware diversity into a superpower—covering our Triton and Inductor extensions for MTIA and the DevX principles that make heterogeneity feel low friction. If you want to see how PyTorch is evolving into the operating system for the AI hardware ecosystem, this is the talk for you.