Introduction#
Meta has recently announced the development of four custom-designed chips specifically for artificial intelligence (AI) tasks. This move is part of the company's broader strategy to expand its data center capabilities.
The MTIA Chip Family#
The new chips belong to the Meta Training and Inference Accelerator (MTIA) family, which the company first introduced in 2023. Meta plans to roll out these four new generations of MTIA chips over the next two years. These chips will enhance tasks such as ranking and recommendations, as well as support advanced generative AI functions, which involve creating content based on user prompts.
Custom Chips for Better Performance#
Meta's Vice President of Engineering, Yee Jiun Song, explained that by designing its own chips, manufactured by Taiwan Semiconductor, the company can achieve better performance and cost efficiency across its data centers. This strategy also diversifies their silicon supply, helping to mitigate the impact of price fluctuations in the market.
Deployment Timeline and Future Plans#
The first of the new chips, the MTIA 300, has already been deployed and is designed to assist in training smaller AI models that enhance user experiences on platforms like Facebook and Instagram. The MTIA 400 is currently being tested and is expected to be operational soon, while the remaining two chips are projected to be in use by 2027. Song noted that releasing new chips every six months is unusual in the tech industry, but it reflects Meta's rapid growth and investment in AI technology.
Conclusion#
Meta's ongoing investment in AI infrastructure includes new data centers in Louisiana, Ohio, and Indiana, with plans to expand further in Texas. The company anticipates that these new chips will have a useful life of over five years, supporting its ambitious AI initiatives.