The finish line was announced on Wednesday that it has developed four novel computer chips that will be used to power the artificial intelligence generative functions and content ranking systems in the tech giant’s own applications. The equipment will become part of Meta’s existing line of chips known as MTIA, or Meta Training and Inference Accelerators.
Meta has partnered with Broadcom to develop the latest open-source RISC-V semiconductors. They are manufactured by Taiwan Semiconductor Manufacturing Corporation, the world’s leading chipmaker.
One of the novel MTIA chips, the MTIA 300, is already in production, and Meta says the other three – MTIA 400, 450 and 500 – are expected to ship in early or slow 2027. Releasing novel silicon so quickly is unusual by most chip industry standards, but virtually unheard of for a social media company that hasn’t historically built its own physical computing infrastructure.
YJ Song, vice president of engineering at Meta, says AI models evolve faster than customary chip development cycles, so AI workloads can change significantly before novel hardware typically reaches production. “Rather than placing bets and waiting for a long time, we are deliberately taking an iterative approach. Each generation of MTIA builds on the previous one, using modular chiplets and incorporating the latest AI workload analytics and hardware technologies,” Song said in blog post.
MTIA 300 will be used primarily to train algorithms that rate and recommend content to the hundreds of millions of people who employ apps like Facebook and Instagram every day. The remaining three chips are designed to handle inference, the process of running trained artificial intelligence models to produce outputs such as text or images.
The MTIA 400, which Meta says delivers performance “competitive with leading commercial products,” has been tested and is expected to hit data centers soon. The MTIA 450 will have twice the amount of high-bandwidth memory as the MTIA 400 and is expected to ship in early 2027. Meta says the MTIA 500, which is expected to arrive slow next year, will have even more memory than the MTIA 450 and will feature “low-precision data innovations.”
The MTIA chips are part of Meta’s broader strategy to marshal as much computing power as possible to develop cutting-edge artificial intelligence. Meta first shared details about its chip development plans in 2023, when it launched its first product under the MTIA banner. As software vendors and AI labs continue to train increasingly powerful AI models, they have begun to announce ambitious plans to build custom chips to meet their own specific AI needs. For example, OpenAI also said cooperates with Broadcom to create custom accelerators, following a path similar to Meta’s.
Earlier this year, Meta has been reported to be decreasing some of its internal efforts to create high-end chips that could compete more directly with leading players like Nvidia. Now, the company appears to be looking to dispel this narrative by announcing a novel roadmap for MTIA chips. However, custom silicon fabrication remains extremely high-priced and technically convoluted, which means Meta will likely continue to purchase most of its AI hardware from other companies, at least for the foreseeable future.
This reality is reflected in the company’s recent chip buying frenzy. Meta unveiled its novel MTIA chips shortly after announcing multi-billion dollar deals with Nvidia and AMD. The social media giant also signed a deal to rent Google-made chips.