MTIA 300: Meta’s First Training Chip with Built-in NICs and Communication-Offloading Engines
Meta's MTIA 300 embeds NICs on-chip, making communication first-class for recommendation model training
“We've made communication a first-class citizen in the chip's design, not an afterthought handled by general-purpose compute cores.”
Meta has revealed MTIA 300, a custom training accelerator that integrates twelve 800 Gbps RDMA NICs directly into the chip package, delivering 1.2 TB/s I/O bandwidth without a PCIe bottleneck. This co-designed hardware-software approach (paired with the HCCL communication library) directly addresses the unique bottleneck of recommendation model training, where embedding tables drive constant AllReduce/AllToAll collectives that starve GPU compute. The announcement signals Meta's deepening vertical integration in AI silicon and a deliberate architectural divergence from general-purpose GPU clusters for its core ranking workloads.