Serving a wide range of AI models on a global scale, while maintaining the lowest possible costs, is one of the most demanding infrastructure challenges...

Meta AI's platform is a resource for artificial intelligence (AI) researchers, machine learning practitioners, and technology enthusiasts. Through articles, research papers, and technical discussions, Meta AI offers insights into AI algorithms, deep learning models, and neural network architectures. Readers can learn about AI applications in various domains, such as computer vision, natural language processing, and reinforcement learning. Additionally, Meta AI provides updates on the latest advancements in AI research, AI ethics, and AI governance to help readers stay informed and engaged in the rapidly evolving field of artificial intelligence.

AI at Meta

Meta details the rapid evolution of its in-house MTIA (Meta Training and Inference Accelerator) chip family, developed with Broadcom. Four generations — MTIA 300, 400, 450, and 500 — were shipped in under two years, with HBM bandwidth increasing 4.5x and compute FLOPS increasing 25x from MTIA 300 to 500. The strategy centers on three pillars: high velocity through modular chiplet architecture enabling ~6-month release cadence, inference-first optimization targeting GenAI workloads, and frictionless adoption via PyTorch-native software stack with vLLM, Triton, and OCP standards. The software stack includes custom compilers built on Torch FX IR, TorchInductor, MLIR, and LLVM, a dedicated collective communications library (HCCL), and a Rust-based user-space driver and firmware. MTIA chips are deployed at scale to power AI experiences for billions of Meta platform users.

Four MTIA Chips in Two Years: Scaling AI Experiences for Billions