Meta Unveils MTIA 300, Its First In-House AI Training Chip with Built-in NICs

Meta disclosed MTIA 300, the first in its family of custom training and inference accelerators designed for ranking and recommendation models¹. The chip integrates two network chiplets containing twelve custom 800 Gbps RDMA NICs, delivering 1.2 TB/s of total I/O bandwidth without crossing a PCIe bus. Sixteen dedicated message engines, each with a RISC-V core and near-memory compute block, handle communication independently, enabling line-rate AllReduce and ReduceScatter collectives without touching the compute grid. Meta reported that running large GEMMs concurrently with collective operations introduces less than 0.5% compute degradation, compared to over 20% on traditional GPUs. On a 150-billion-parameter production recommendation model across 40 accelerators, MTIA 300 communication time was 3.9 times faster than an equivalent GPU cluster.