Meta Platforms has recently shared new information about its data center projects, which aim to enhance support for artificial intelligence (AI) work. One notable development is the company’s in-house custom chip “family” being developed to bolster their AI capabilities.
In a series of blog posts, Meta, the parent company of Facebook and Instagram, revealed that it had designed a first-generation chip in 2020 as part of its Meta Training and Inference Accelerator (MTIA) program. The primary objective was to enhance the efficiency of recommendation models utilized for delivering ads and other content in users’ news feeds.
Earlier reports from Reuters had suggested that Meta had no plans to widely deploy its initial in-house AI chip and was already working on a successor. The blog posts, however, described the first MTIA chip as a valuable learning experience.
The first iteration of the MTIA chip primarily focused on a process called inference in AI, where algorithms, trained on vast amounts of data, make decisions about whether to display, for example, a dance video or a cat meme as the next post in a user’s feed. According to the blog posts, Joel Coburn, a software engineer at Meta, explained during a presentation that while Meta initially relied on graphics processing units (GPUs) for inference tasks, they were not well-suited for this work. Coburn stated that GPUs were inefficient for real models, despite significant software optimizations, making them challenging and expensive to deploy effectively. This is why the development of MTIA was deemed necessary.
Meta declined to provide specific timelines for deploying the new chip or further details on plans to create chips capable of both training and inference tasks.
Over the past year, Meta has undertaken a substantial project to upgrade its AI infrastructure. Company executives realized that they lacked the necessary hardware and software to meet the demands from product teams developing AI-powered features. Consequently, Meta scrapped its plans for a large-scale rollout of an in-house inference chip and shifted focus to a more ambitious chip capable of handling both training and inference, as reported by Reuters.
The blog posts by Meta acknowledged that their first MTIA chip encountered challenges with high-complexity AI models. However, they emphasized that the chip outperformed competitor chips in terms of efficiency when dealing with low- and medium-complexity models. Furthermore, Meta stated that the MTIA chip consumed only 25 watts of power, significantly less than leading chips from suppliers like Nvidia, and employed an open-source chip architecture called RISC-V.
In addition to chip development, Meta provided an update on their plans to revamp their data centers with modern AI-focused networking and cooling systems. The company aims to commence construction on their first facility of this kind later this year. The redesigned data centers are projected to be 31 percent cheaper and can be built in half the time required for Meta’s current data centers, as explained by an employee in a video outlining the changes.
Lastly, Meta mentioned their AI-powered system designed to assist engineers in creating computer code. Similar tools have been offered by Microsoft, Amazon.com, and Alphabet.
