Meta Platforms revealed new initiatives on Thursday aimed at enhancing its data centers for artificial intelligence applications, including the development of a custom chip “family” in-house.
The parent company of Facebook and Instagram outlined these efforts in a series of blog posts, noting that it designed a first-generation chip in 2020 under the Meta Training and Inference Accelerator (MTIA) program. This initiative aims to boost the efficiency of its recommendation algorithms that deliver ads and other content in users’ news feeds.
Earlier reports from Reuters indicated that Meta would not be widely deploying this initial AI chip and was already focused on a successor. In the recent blog entries, the company described the first MTIA chip as a valuable learning opportunity.
“From this initial program, we have learned invaluable lessons that we are incorporating into our roadmap,” the company stated.
The initial MTIA chip concentrated solely on a process known as inference, where algorithms analyze vast amounts of data to decide which content, such as a dance video or a cat meme, should be displayed next in a user’s feed.
A representative for Meta did not provide specifics regarding deployment timelines or elaborate on plans for chips capable of model training.
This past year, Meta undertook a significant upgrade of its AI infrastructure after company executives recognized the inadequacy of its hardware and software in meeting the needs of product teams creating AI-driven features.
In line with these upgrades, the company abandoned plans for a large-scale rollout of an inference chip and is now working on a more advanced design that can perform both training and inference tasks, according to Reuters.
Meta acknowledged in its blog posts that its first MTIA chip struggled with high-complexity AI models but noted that it performed more effectively than competitor chips with lower and medium-complexity models.
This chip was reported to consume only 25 watts of power, significantly less than leading chips from companies like Nvidia. It was designed using an open-source architecture known as RISC-V, according to Meta.
In addition to discussing its chip developments, Meta offered insights into its plans to redesign its data centers to integrate modern AI-centric networking and cooling systems, with intentions to commence construction on its first such facility this year.
The new data center design is projected to be 31 percent less expensive and capable of construction in half the time compared to the current facilities, as noted by an employee in a video detailing the changes.
© Thomson Reuters 2023
