Meta Debuts MTIA 300-500 AI Chips to Cut Costs, Boost Performance
Meta introduced four new MTIA AI chips—300, 400, 450 and 500—designed to handle ranking, recommendations and high-end inferencing with shared infrastructure for easy upgrades. The MTIA 400 claims raw performance competitive with leading GPUs and cost savings, and Meta plans full deployment by 2027.
1. Meta's MTIA Chip Family
Meta unveiled four custom AI processors—MTIA 300, 400, 450 and 500—designed for internal training and inference workloads. These chips share common infrastructure to enable straightforward upgrades as performance needs evolve.
2. MTIA 400 Targets Generative AI and Recommendations
The MTIA 400 can handle generative AI tasks alongside ranking and recommendation models and supports rack configurations of up to 72 units. Meta claims it matches leading commercial GPUs in raw performance while delivering cost efficiencies.
3. Enhanced Memory in MTIA 450 and 500
Building on the MTIA 400, the MTIA 450 adds faster high-bandwidth memory and the MTIA 500 increases capacity with even faster speeds. This tiered approach allows Meta to allocate resources based on workload intensity.
4. Deployment Timeline and Strategic Goals
Meta has already integrated select MTIA chips and plans full deployment across its data centers by 2027. Developing in-house processors aims to reduce reliance on external vendors and control AI infrastructure costs.