Meta is reportedly testing its first in-house AI training chip

Breaking: A Big Tech company is ramping up its AI development. (Whaaat??) In this case, the protagonist of this now-familiar tale is Meta, which Reuters reports is testing its first in-house chip for AI training. The idea is to lower its gargantuan infrastructure costs and reduce its dependence on NVIDIA (a company that apparently brings out Mark Zuckerberg’s “adult language” side). If all goes well, Meta hopes to use it for training by 2026.

Meta has reportedly kicked off a small-scale deployment of the dedicated accelerator chip, which is designed to specialize in AI tasks (and is, therefore, more power-efficient than general-purpose NVIDIA GPUs). The deployment began after the company completed its first “tape-out,” the phase in silicon development where a complete design is sent for a manufacturing test run.

The chip is part of the Meta Training and Inference Accelerator (MTIA) series, the company’s family of custom in-house silicon focused on generative AI, recommendation systems and advanced research.

Last year, the company started using an MTIA chip for inference, a predictive process that happens behind the scenes in AI models. Meta began using the inference one for its Facebook and Instagram news feed recommendation systems. Reuters reports that it plans to start using the training silicon for that as well. The long-term plan for both chips is said to begin with recommendations and eventually use them for generative products like the Meta AI chatbot.

The company is one of NVIDIA’s biggest customers after placing orders for billions of dollars’ worth of GPUs in 2022. That was a pivot for Meta after it bailed on a previous in-house inference silicon that failed a small-scale test deployment — much like the one it’s doing now for the training chip.