MTIA 300: Meta’s First Training Chip with Built-in NICs and Communication-Offloading Engines

Meta's MTIA 300 training chip gives communication its own hardware so that exchanging data need not consume the same compute resources as training. The chip account describes built-in network interfaces and dedicated message engines that execute coordinated transfers between accelerators. Its communication library prepares dependent operations for those engines to run after dispatch.

On a 150-billion-parameter recommendation model across 40 accelerators, the company reports a 3.9-fold communication speedup over an equivalent GPU cluster. That result measures the data-exchange portion of this workload; the architectural contribution is separating that work from the chip's training computation.