Meta has introduced four new generations of its in-house AI accelerator chips as the company ramps up efforts to power generative AI services across its platforms used by billions of people.
The company said the new chips—MTIA 300, MTIA 400, MTIA 450 and MTIA 500—are part of its Meta Training and Inference Accelerator (MTIA) family and will be deployed across its data centres between 2026 and 2027.
The chips are built to handle a wide range of AI workloads, including recommendation systems, generative AI training and large-scale inference.
“Serving a wide range of AI models on a global scale, while maintaining the lowest possible costs, is one of the most demanding infrastructure challenges in the industry,” Meta said in a blog post. “MTIA plays an important role in cost-effectively powering AI experiences for the billions of people who use Meta’s products.”
Meta said it has already deployed hundreds of thousands of earlier MTIA chips in production and tested the architecture with large language models such as Llama.
The company developed the chips in partnership with semiconductor firm Broadcom as part of a broader strategy to reduce reliance on external hardware and optimise infrastructure for its AI workloads.
The MTIA 300 chip is currently used for training ranking and recommendation models, which historically formed the bulk of Meta’s AI workloads. Building on that foundation, the MTIA 400 introduces higher compute performance and increased memory bandwidth, allowing it to support both recommendation systems and generative AI models. Meta said the chip delivers 400% higher FP8 compute performance compared with MTIA 300.
Later versions are more specialised for generative AI inference. MTIA 450 doubles high-bandwidth memory (HBM) bandwidth compared with MTIA 400 and adds hardware optimisations for attention and feed-forward network operations used in modern AI models.
The upcoming MTIA 500 further expands memory bandwidth and capacity while improving compute efficiency, making more cost-effective AI inference at scale possible.
Both of these chips are expected to be rolled out through 2027.
Across the MTIA 300 to 500 generations, Meta said HBM bandwidth has increased by 4.5 times and compute performance by roughly 25 times within two years.
The company attributes the rapid development cycle to a modular chiplet-based design that allows components to be upgraded independently. Meta said this approach enables it to release new chips roughly every six months.
Meta said its MTIA strategy is built around three key principles: rapidly developing new chip generations, prioritising AI inference workloads, and ensuring easy adoption by designing the chips around widely used industry standards such as PyTorch.
The company added that the chips are made to integrate seamlessly with widely used AI tools such as PyTorch, vLLM and Triton, allowing developers to deploy models without significant changes to existing workflows.
Meta expects that new MTIA chips will help it run generative AI systems more efficiently and handle the growing demand for AI features across its apps.
ALSO READ: Zoom Expands Enterprise Agentic AI Platform
