Meta Platforms on Thursday shared new details on projects it was pursuing to make its data centers better suited to supporting artificial intelligence work, including a custom chip "family" that it said it was developing in-house.
The Facebook and Instagram owner said in a series of blog
posts that it designed a first-generation chip in 2020 as part of the Meta
Training and Inference Accelerator (MTIA) program, which was aimed at improving
efficiency for the recommendations models it uses to serve ads and other
content in news feeds.
Reuters previously reported that the company was not
planning to deploy its first in-house AI chip widely and was already working on
a successor. The blog posts portrayed the first MTIA chip as a learning
opportunity.
"From this initial program, we have learned invaluable
lessons that we are incorporating into our roadmap," it wrote.
The first MTIA chip was focused exclusively on an AI process
called inference, in which algorithms trained on huge amounts of data make
judgments about whether to show, say, a dance video or a cat meme as the next
post in a user's feed, the posts said.
A Meta spokesperson declined to comment on deployment
timelines or elaborate on the company's plans to develop chips that could train
the models as well.
Meta has been engaged in a massive project to upgrade its AI
infrastructure this past year, after executives realized it lacked the hardware
and software needed to support demand from product teams building AI-powered
features.
As part of that, the company scrapped plans for a
large-scale rollout of an in-house inference chip and started work on a more
ambitious chip capable of performing both training and inference, according to
the Reuters reporting.
Meta acknowledged in its blog posts that its first MTIA chip
stumbled with high-complexity AI models, although it said the chip handled low-
and medium-complexity models more efficiently than competitor chips.
The MTIA chip also used only 25 watts of power — a fraction
of what market-leading chips from suppliers such as Nvidia consume — and used
an open-source chip architecture called RISC-V, Meta said.
In addition to detailing its chip work, Meta provided an
update on plans to redesign its data centers around more modern AI-oriented
networking and cooling systems, saying it would break ground on its first such
facility this year.
The new design would be 31 percent cheaper and could be
built twice as quickly as the company's current data centers, an employee said
in a video explaining the changes. -Reuters
0 comments:
Post a Comment