Meta has built networking and collective-communication hardware directly into MTIA 300, its in-house accelerator for training ranking and recommendation models. The design targets workloads where moving data among hundreds of chips can matter as much as raw computing speed.
Recommendation systems rely heavily on embedding tables, which Meta says can hold more than 99% of a model’s parameters. MTIA 300 includes two network chiplets carrying 12 custom 800-gigabit-per-second network interfaces, for 1.2 terabytes per second of total input and output bandwidth. Sixteen dedicated message engines handle communication separately from the chip’s main compute grid.
Meta reports that large matrix calculations and network operations can overlap with less than 0.5% degradation in compute throughput. Its comparison GPU lost more than 20% when both ran together, though those figures come from Meta’s own testing and do not establish performance across all workloads.
The chip was co-designed with HCCL, Meta’s communication software. By treating the network as part of the accelerator rather than an external attachment, Meta is extending its custom-silicon strategy from computation into the data movement that increasingly constrains large recommendation systems.