Hardware
Meta Unveils MTIA AI Chips for Scalable AI Infrastructure
Meta has deployed hundreds of thousands of MTIA chips to power AI experiences for billions of users, with new generations set for 2026 and 2027.
Image: Meta AI
Meta has announced the continued development and deployment of its MTIA family of AI chips, designed to scale AI experiences for billions of users. The company has deployed hundreds of thousands of MTIA chips in production, supporting a wide range of AI models across Meta’s platforms. The MTIA family includes four generations of chips, with MTIA 300, 400, 450, and 500 each optimized for different AI workloads, from recommendation systems to large language models. The latest chips are expected to be deployed in 2026 and 2027, reflecting Meta’s iterative approach to chip development. This strategy allows for rapid adaptation to evolving AI models and workload demands, ensuring cost-effective and high-performance solutions for global AI infrastructure. The MTIA chips are developed in close partnership with Broadcom and are part of Meta’s broader AI infrastructure strategy. *Source: [metaai](https://ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions/)*
Key points
- Meta has deployed hundreds of thousands of MTIA chips in production to support AI models across its platforms.
- MTIA 300 is optimized for R&R training and includes built-in NIC chiplets and near-memory compute.
- MTIA 400 delivers 400% higher FP8 FLOPS and 51% higher HBM bandwidth compared to MTIA 300.
- MTIA 450 doubles HBM bandwidth from MTIA 400 and introduces low-precision data types for GenAI inference.
- MTIA 500 increases HBM bandwidth by 50% and MX4 FLOPS by 43% compared to MTIA 450.
- From MTIA 300 to MTIA 500, HBM bandwidth increases by 4.5x and compute FLOPS by 25x.
- MTIA 450 delivers 6x the MX4 FLOPS of FP16/BF16 and supports mixed low-precision computation.