Inside the hardware and co-design of MTIA

Meta Details the Hardware and Co-Design Behind Its Second-Gen MTIA

At AI Infra @ Scale 2024, Meta engineers Joel Colburn, Junqiang Lan, and Jack Montgomery presented the company’s work on the second generation of MTIA, its in-house accelerator for training and inference.

The presentation covers the co-design process behind Meta’s first custom silicon for AI workloads, emphasizing the PyTorch software stack and the model architectures powering key Meta applications. The team demonstrates how MTIA delivers the performance, efficiency, and developer experience required to push production models live. Several co-design examples are highlighted, showing how dedicated silicon features accelerate Meta’s specific models in practice.