Meta has initiated trials for its first custom-designed chipsets intended for training artificial intelligence (AI) models. Reports indicate that the company has rolled out a small batch of these processors to evaluate their performance and efficiency. Depending on the outcome of these assessments, Meta plans to move forward with large-scale production of the hardware, which is part of its Meta Training and Inference Accelerator (MTIA) series.
As detailed in a report by Reuters, these AI chipsets were developed in partnership with Taiwan Semiconductor Manufacturing Company (TSMC). Meta has recently completed the final phase of chip design, known as tape-out, and has now commenced limited deployment of the technology.
This marks not the first attempt by Meta to produce AI-centric chipsets; last year, the company introduced Inference Accelerators tailored for AI inference tasks. However, until this point, Meta lacked proprietary hardware accelerators necessary for training its Llama suite of large language models (LLMs).
Sources within the company suggest that Meta’s overarching goal in creating these in-house chipsets is to reduce the costs associated with the infrastructure required to operate complex AI systems. This initiative is aimed at supporting internal projects, consumer-oriented products, and tools for developers.
In an interesting development from January, Meta CEO Mark Zuckerberg announced the completion of the Mesa Data Center expansion in Arizona, which has since become operational. The deployment of the new training chipsets is expected to take place at this facility.
The report also indicated that the initial application of the new chipsets will be within Meta’s recommendation engine, which drives content across its various social media platforms. The use cases are anticipated to broaden in the future to include generative AI solutions such as Meta AI.
During the same month, Zuckerberg shared in a Facebook post that the company intends to allocate up to $65 billion (approximately ₹5,61,908 crore) towards AI initiatives by 2025. This investment will encompass the Mesa Data Center expansion and the recruitment of additional personnel to bolster its AI teams.
