Meta is Creating Four New Chips to Enhance Its AI and Recommendation Technologies.

On Wednesday, Meta announced the creation of four new computer chips designed to enhance generative AI features and content ranking systems within its applications. These chips will expand upon Meta’s existing MTIA line, which stands for Meta Training and Inference Accelerators.

In collaboration with Broadcom, Meta has developed these latest semiconductors based on the open-source RISC-V architecture. They are being manufactured by Taiwan Semiconductor Manufacturing Corporation, the leading chip producer globally.

The MTIA 300 chip is already in production, while the other three models—the MTIA 400, 450, and 500—are projected to be shipped between early and late 2027. Launching new silicon at such a pace is rare within the chip industry, particularly for a social media company that has historically refrained from creating its own physical computational infrastructure.

YJ Song, a Meta engineering vice president, noted that AI models are advancing more rapidly than traditional chip development timelines, which means AI workloads could greatly change when new hardware typically becomes available. “Instead of making a long-term commitment and waiting, we consciously take an iterative approach. Each MTIA generation builds on the previous one using modular chiplets and the latest insights into AI workloads and hardware technologies,” Song mentioned in a blog post.

The MTIA 300 is primarily intended for training algorithms that rank and recommend content to the millions who engage with apps like Facebook and Instagram daily. The remaining three chips aim to support inference, the task of executing trained AI models to generate outputs such as text or images.

The MTIA 400, which Meta asserts offers performance “on par with leading commercial products,” has undergone testing and is expected to be deployed in data centers soon. The MTIA 450 is designed to include double the high-bandwidth memory of the MTIA 400, with shipments anticipated in early 2027. According to Meta, the MTIA 500, scheduled to be released later next year, will boast even more memory than the MTIA 450 and feature “innovations in low-precision data.”

These MTIA chips are integral to Meta’s overarching strategy to accumulate ample computing power for pioneering artificial intelligence advancements. The company first disclosed details about its chip development intentions in 2023 with the launch of its initial MTIA product. As software firms and AI laboratories train increasingly powerful models, they have started announcing ambitious plans for custom chips tailored to their unique AI requirements. OpenAI, for instance, has also indicated a partnership with Broadcom to create customized accelerators, mirroring Meta’s approach.

Earlier this year, reports indicated that Meta was scaling back its in-house efforts to develop high-end chips to compete directly with leading companies like Nvidia. However, the company now seems eager to counter that impression by unveiling this new roadmap for MTIA chips. Despite this, the process of creating custom silicon remains highly expensive and technically intricate, leading Meta to likely keep purchasing the majority of its AI hardware from external firms, at least for the foreseeable future.

This situation is reflected in Meta’s recent acquisition of chips. The introduction of its new MTIA chips followed the announcement of multibillion-dollar partnerships with Nvidia and AMD, along with an agreement to utilize chips manufactured by Google.