On this discuss from AI Infra @ Scale 2024, Joel Colburn, a software program engineer at Meta, technical lead Junqiang Lan, and software program engineer Jack Montgomery talk about the second generation of MTIA, Meta’s in-house coaching and inference accelerator.
They cowl the co-design course of behind constructing the second era of Meta’s first-ever customized silicon for AI workloads, together with the PyTorch software program ecosystem, and the mannequin architectures for Meta’s key purposes. They show how MTIA achieves the efficiency, effectivity, and developer expertise to efficiently launch fashions into manufacturing. Additionally they spotlight a number of co-design examples the place particular silicon options are utilized to speed up Meta’s fashions.