ExclusiveMeta To Deploy Inhouse Custom Chips This Year To Power AI Drive Memo
Katie Paul, Stephen Nellis and Max A. Cherney
(Reuters) - Facebook owner Meta Platforms plans to launch a new version this year to support its push for artificial intelligence (AI) in data centers, according to an internal document seen by Reuters on Thursday.
The chip, the second generation of Meta's in-house silicon line announced last year, will help Meta reduce its reliance on existing NVIDIA chips and control rising costs in the race to launch AI products.
The world's largest social media company is struggling to leverage computing power for its power-hungry artificial intelligence products on Facebook, Instagram and WhatsApp, as well as hardware devices like Ray-Ban smart glasses. the armory Reconfigured to accommodate custom chips and data centers
According to Dylan Patel, the team's founder, in a study by SemiAnalysis Silicon, at a meta scale, successful deployment of the chip could reduce annual energy costs by tens of millions of dollars and chip acquisition costs by tens of millions of dollars.
The chips, infrastructure and power needed to power AI applications have become a major investment drain for tech companies, partially offsetting the hype surrounding the technology.
A Meta spokesperson confirmed that the upgraded chip, which will come with several thousand graphics processing units (GPUs) (AI chips) that the company has already purchased, plans to go into production in 2024.
"We find our built-in accelerators very helpful with commercially available GPUs to provide the best combination of performance and efficiency in meta-intensive workloads," the spokesperson said.
Meta CEO Mark Zuckerberg said last month that the company expects to receive about 350,000 Flaga "H100" processors from Nvidia by the end of the year, the most in-demand GPUs used in artificial intelligence. Along with other vendors, Meta says it will muster a total of 600,000 computing power equivalent to the H100.
Expanding the chip as part of this plan is a positive for Mater's in-house AI silicon project following its decision to launch the first iteration of the chip in 2022.
Instead, the company chose to buy billions of dollars worth of GPUs with a monopoly on an artificial intelligence process called training, which involves feeding a model a huge set of data to teach it how to perform a task.
The new chip, internally called Artemis, performs a similar process known as inference, where models use their own algorithms to judge categories and answer user questions.
Reuters reported last year that Meta was working on a giant chip that could perform GPU-like training and insight.
The Menlo Park, California-based company shared details about the first batch of its Meta Training and Results Accelerator (MTIA) program last year. The ad presented that version of the chip as a learning opportunity.
Despite this initial hurdle, the inference chip could be more efficient at reducing meta-recommendations than power-hungry Nvidia processors, Patel said.
"A lot of money can be spent and labor can be saved," he said.
(Reporting by Katie Paul and Stephen Nellis in New York and Max A. Charney in San Francisco; Editing by Kenneth Lee and Mark Porter)
