In order to further assist artificial intelligence work, Meta Platforms revealed additional information on its data center initiatives on Thursday. One of the specifics was a proprietary chip "family" that is being created in-house.
In a series of blog articles, the owner of Facebook and Instagram said that as part of the Meta Training and Inference Accelerator (MTIA) program, it will design a first-generation microprocessor in 2020. The goal was to make the recommendation models, which are used to distribute adverts and other material in news feeds, more effective.
According to a previous report by Reuters, the company had no intentions of extensively implementing its initial in-house AI chip and was already in the process of developing a follow-up chip. The blog posts depicted the first MTIA chip as a valuable learning experience.
As mentioned in the posts, the primary focus of the first MTIA chip was on an AI process known as inference. In this process, algorithms trained on extensive data are utilized to make decisions on whether to display content like a dance video or a cat meme as the next post in a user's feed.
During a presentation on the new chip, Joel Coburn, a software engineer at Meta, explained that Meta initially turned to graphics processing units (GPUs) for inference tasks but discovered they were not optimal for such work.
Coburn mentioned that the efficiency of GPUs for real models, despite substantial software optimizations, is comparatively low, resulting in difficulties and high expenses when it comes to practical deployment. He emphasized that this is precisely why MTIA is needed.
A Meta representative declined to comment on the new chip's deployment schedule or go into further detail about the company's intentions to create chips that could also train the models.
Since executives recognized it lacked the technology and software to accommodate demand from product teams developing AI-powered services, Meta has been working on a significant effort to update its AI infrastructure.
As a result, the business abandoned plans for a wide-scale rollout of an internal inference chip and began developing a more ambitious chip that could conduct both training and inference, according to some media.
In Meta's blog posts, they acknowledged that their initial MTIA chip encountered challenges when dealing with high-complexity AI models. However, they also highlighted that the chip demonstrated more efficient performance than competitor chips when handling low- and medium-complexity models.
Meta stated that the MTIA chip consumed only 25 watts of power, which is significantly less compared to the power consumption of market-leading chips from suppliers like Nvidia Corp. Additionally, Meta mentioned that the MTIA chip utilized an open-source chip architecture called RISC-V.
Meta also shared an update on its intentions to revamp its data centers with modern AI-oriented networking and cooling systems. They revealed plans to commence construction on their first facility of this kind within the current year.
In a video explaining the modifications, an employee mentioned that the new design would be 31% more cost-effective and could be constructed in half the time compared to Meta's existing data centers.
To assist its engineers in writing computer code, Meta claimed to have a system driven by artificial intelligence (AI), similar to those provided by Alphabet Inc., Amazon.com Inc., and Microsoft Corp.