Meta Platforms has unveiled its MTIA 300 custom AI accelerator, which integrates two network chiplets with six 800 Gb/s RDMA NICs each, delivering 1.2 TB/s of I/O bandwidth and eliminating CPU-mediate…
Meta has detailed MTIA 300, its first in-house accelerator optimized for training ranking and recommendation models, integrating networking and collective communication directly into the chip to addre…
Meta Platforms Inc. presented its Meta Training and Inference Accelerators (MTIA) custom AI silicon at Hot Chips 2026, detailing a roadmap of four generations (MTIA 300, 400, 450, 500) to reduce relia…
Meta Platforms Inc. unveiled MTIA 300, the first chip in its MTIA family designed for both training and inference, optimized for ranking and recommendation models, and featuring built-in NIC chiplets …
Meta Platforms Inc. unveiled four new generations of its in-house Meta Training and Inference Accelerator (MTIA) chips on March 11, 2026, and expanded a partnership with Broadcom Inc. to co-develop th…
Meta's HCCL collective communication library, co-designed with its MTIA 300 accelerator, achieves up to 940 GB/s on intra-rack collectives for training while introducing less than 0.5% degradation to …
Meta plans to manufacture its next-generation Iris AI chip in September, targeting a 25x increase in compute FLOPS and a 4.5x improvement in memory bandwidth by 2027. The in-house silicon push, fabric…