Meta ramps up AI efforts by building chip arsenal, consolidating teams

0
112
Meta Connect conference in Menlo Park

NEW YORK (Reuters) -Meta CEO Mark Zuckerberg said on Thursday that the company was bringing its AI research team “closer together” with a more business-focused generative AI team launched last year, doubling down on a push to get the technology into products.

The social media giant was building out its infrastructure to accommodate the push and planned to have about 350,000 H100 GPUs (graphics processing units) from chip designer Nvidia by the end of the year, Zuckerberg said in posts on Meta’s Instagram and Threads platforms.

In combination with equivalent chips from other suppliers, Meta will have about 600,000 total GPUs by the end of the year, he said.

That amount would make Meta’s system, once finished, among the largest in the technology industry.

By contrast, Amazon last autumn said it was building out a system with 100,000 of its Trainium2 chips, while Oracle brought online a system with 32,000 Nvidia H100 GPUs.

A Meta spokesperson declined to specify which GPU suppliers the company was using aside from market leader Nvidia, though it has said publicly it also plans to use chips from AMD. Reuters has previously reported it has an internally designed GPU-like chip in the works, as well.

Meta has scrambled to build out its computing arsenal to support a mobilization around generative AI this past year, after years of producing leading research on the technology through its FAIR team but scant focus on building it into its core social media products and AR/VR hardware devices.

The company convened a “GenAI” team last year to lead an effort to change that after the breakout success of OpenAI’s ChatGPT chatbot in late 2022.

Since then, Meta has launched a commercial version of its Llama large language model, ad tools that can generate image backgrounds from text prompts and a “Meta AI” chatbot that can be accessed directly via its Ray-Ban smart glasses.

In his posts on Thursday, Zuckerberg said the company was currently training a third version of the Llama model.

He also tied the AI investments back to the pivot to the AR/VR-driven metaverse vision that inspired him to change the company’s name to Meta in 2021, saying people would “need new devices” like glasses for interacting with AI.

(Reporting by Katie Paul in New York; additional reporting by Stephen Nellis in San Francisco; Editing by Franklin Paul and Lisa Shumaker)

tagreuters.com2024binary_LYNXMPEK0H0TP-VIEWIMAGE