Meta Unveils Four Custom AI Chips

Meta announced on Wednesday that it is developing four new generations of custom AI chips within the next two years to support its artificial intelligence workloads.

The specialized chips are part of the Meta Training and Inference Accelerator family, which the company first revealed publicly in 2023 before releasing a second-generation version in 2024. The four new chips, identified as MTIA 300, MTIA 400, MTIA 450, and MTIA 500, are designed to enhance various aspects of Meta's AI operations, including ranking and recommendations systems as well as advanced generative AI tasks.

The MTIA 300 was deployed a few weeks ago and is intended to help train smaller AI models that support Meta's core ranking and recommendation tasks, including showing relevant content and online ads within Facebook and Instagram. The MTIA 400, which Meta has finished testing, is designed for generative AI and ranking tasks and can be interconnected within a server framework housing up to 72 chips.

Meta asserts that the MTIA 400 is its first chip offering both cost efficiency and raw performance that competes with leading commercial products, though the company did not specify which products it is comparing against. The MTIA 450 processor enhances the MTIA 400 capabilities with faster high-bandwidth memory, while the MTIA 500 adds more memory capacity with even greater speed.

The company stated it is releasing chips at an unusually fast pace, approximately one new chip every six months, with all four generations expected to be operational by the end of 2027. Albert Song, head of Meta's custom silicon efforts, explained that Meta finds itself building out capacity quickly and spending heavily on capital expenditures, making it essential to deploy state-of-the-art chips at any given time.

A significant advantage for Meta is that all MTIA chips share a common foundational infrastructure at the system level, including the same chassis, rack, and network infrastructure. This modular design allows each new chip generation to slot into the same physical footprint, accelerating the transition from silicon to production deployment.

Meta has also recently entered into multiyear, multigenerational chip supply agreements with both Nvidia and AMD, indicating it plans to maintain a diversified hardware strategy.