Meta has announced four new generations of its MTIA training and inference accelerator chips.
With demand for AI chips rising and supplies tightening, Meta is taking its AI computing needs into its own hands and developing more of its own chips: It will produce four new generations of chips over the next two years.
Cloud computing giants including Meta, AWS, and Google have been keen to develop their own chips to improve the performance of their own data centers. Meta started its own chip program in 2023, when it implemented the Meta Training and Inference Accelerator (MTIA), a family of custom-built silicon chips to power its AI workloads efficiently.
The MTIA 300, which Meta will use for ranking and recommendations training, is already in production, Meta said. It will use the other planned chips, the MTIA 400, 450, and 500, mainly for generative AI inference production, it said.
Plowing its own AI furrow is becoming a thing for the owner of Facebook: Although it was an early supporter of open-source large language models, it has shunned an industry group promoting open models and plans to develop proprietary models that it can charge for.
However, Meta is not completely striking out on its own, it is still buying chips from AMD and from ARM.
In addition to designing its own chips, Meta is also looking at other ways to more closely integrate the various strands of its business. In January, it said it will bring together network and AI data center development under a single umbrella.
