Meta has announced the development of an AI training and inference chip

Meta Platforms has released new information about its data center projects aimed at enhancing support for artificial intelligence (AI) work. The company, which owns Facebook and Instagram, disclosed in a series of blog posts that it has been developing a custom chip “family” in-house. The initial chip, developed in 2020 under the Meta Training and Inference Accelerator (MTIA) program, focused on improving efficiency for recommendation models used in serving ads and content on news feeds. While Meta has not provided specific deployment timelines, the blog posts highlighted the learning experience gained from the first MTIA chip. The company found that graphics processing units (GPUs) were not well-suited for inference tasks, leading to the development of the MTIA chip, which offers higher efficiency and lower power consumption. Meta has been undertaking a major AI infrastructure upgrade to meet the demands of its product teams, resulting in the abandonment of plans for a large-scale rollout of an in-house inference chip and a shift towards a more ambitious chip capable of both training and inference. The MTIA chip demonstrated efficient performance for low- and medium-complexity models, consuming only a fraction of the power compared to leading supplier chips like Nvidia. Meta also shared plans to redesign its data centers, incorporating AI-oriented networking and cooling systems, which are expected to be cheaper and faster to build compared to existing facilities. Furthermore, Meta mentioned its AI-powered system to assist engineers in coding, similar to tools offered by Microsoft, Amazon.com, and Alphabet.