· Updated
MTIA) chip family, developed with Broadcom, accelerating to four chip generations — MTIA 300, 400, 450, and 500 — in under two years. The strategy uses modular chiplets that upgrade separately, with MTIA 400, 450, and 500 sharing the same chassis and rack infrastructure.From MTIA 300 to MTIA 500, HBM bandwidth increases 4.5x and compute FLOPS increase 25x (from MX8 to MX4 low-precision formats). MTIA 450 doubles HBM bandwidth over MTIA 400 for GenAI inference decode and introduces hardware acceleration for attention and FFN computation. MTIA 500 adds 50% more HBM bandwidth and up to 80% more HBM capacity. Both are scheduled for mass deployment in 2027.
Review the published chip specs and deployment roadmap to see how MTIA 450 and 500 compare to leading commercial GPU products for GenAI inference workloads.
Custom silicon is critical to scaling next-gen AI. We’re detailing the evolution of the Meta Training and Inference Accelerator (MTIA), our homegrown silicon family designed to power the next era of AI experiences. Traditional chip cycles span years, but model architectures change in months. To close this gap, we’ve accelerated MTIA development to release four generations in just two years. See our roadmap and tech specs here: https://t.co/ZGvo2TQAR2
More like this
MetaFeb 24
NVIDIAMar 19
AnthropicApr 20