Meta reports that on a 150-billion-parameter production recommendation model running across 40 accelerators, MTIA 300's total communication time was 3.9 times faster than an equivalent GPU cluster. MTIA 300 is the first of Meta's in-house accelerators optimized for training ranking and recommendation models, with network chiplets integrated into the chip package. Meta co-designed the chip with HCCL, its communication library, which compiles collectives into subgraphs executed autonomously by dedicated message engines.
No score is assigned. Sources and their independence are shown in the citation chain below.