Meta has begun testing its own AI chips for training neural networks

Back in May of the year before last, Meta✴ Platforms shared details about its own MTIA family of chips, which it will use to speed up its artificial intelligence systems. According to Reuters, training of the company’s AI systems using these chips has already begun.

Image source: Meta✴ Platforms

For now, training is being carried out using a small number of chips of its own design, but if successful, Meta✴ intends to launch these chips into mass production. It is expected that the transition to components of its own design will allow Meta✴ to reduce the costs of maintaining its computing infrastructure. This year, the company’s expenses will range from $114 to $119 billion, of which $65 billion will be spent on infrastructure development.

Meta✴’s own chips are expected to reduce energy costs when performing calculations related to the operation of artificial intelligence systems. Specialized solutions in this area may be more efficient than general-purpose GPUs. The Taiwanese company TSMC is expected to produce chips for Meta✴. According to sources, the preparation for production and the release of a pilot batch of chips could cost Meta✴ hundreds of millions of US dollars. If the first batch does not meet the developers’ expectations, a new stage of development will require comparable costs and several months of waiting. According to rumors, Meta✴ already had to cancel the release of a chip for reasoning AI at previous stages of development and redirect funds to purchase a large number of Nvidia accelerators. Meta✴ Platforms remains one of the largest buyers of Nvidia accelerators to this day. It does not intend to completely abandon their use.

Last year, Meta✴ began experimenting with using its chips to work with reasoning AI, as applied to its social networks. Meta✴ officials said last week that the company will start using its own chips to train language models starting in 2026. They will first find application in conjunction with recommendation algorithms, and will later be adapted for generative artificial intelligence. Meta✴ management describes the company’s progress in developing AI chips as not very stable, but considers the first generation of solutions for reasoning artificial intelligence to be very successful.