Meta plans to use a new version of its proprietary chip to support artificial intelligence efforts in its data centers this year, according to an internal company document seen by Reuters.
The chip could help reduce the company's dependence on market-dominant Nvidia chips and control the high costs associated with using artificial intelligence in the race to launch technology products. Amnesty International.
The new chip, internally called Artemis, is the second generation of Meta's in-house silicon announced last year.
Meta is increasing the processing power of its generative AI products, which it adds to apps like Facebook, Instagram, and WhatsApp, as well as devices like Ray-Ban smart glasses.
Meta is spending billions of dollars to build its own line of chips and reconfigure its data centers accordingly.
Depending on Meta's scale, successful deployment of new chips could save hundreds of millions of dollars in energy costs and billions of dollars in chip procurement costs each year.
The chips, infrastructure, and power needed to run AI applications have become a huge hole, absorbing investments by technology companies and erasing some of the gains made during the technology boom.
A Meta spokesperson confirmed that the company plans to put the updated chip into production in 2024, and said it will work with hundreds of thousands of proprietary, commercially available AI GPUs that the company has purchased.
“We believe our internally developed accelerators are an excellent complement to commercial GPUs and can deliver the best combination of performance and efficiency in meta-data workloads,” the spokesperson said in a statement.
CEO Mark Zuckerberg said last month that Meta plans to buy about 350,000 H100 chips from Nvidia by the end of the year. The Meta combines computing power equivalent to a total of 600,000 H100 chips.
The use of its chips represents a positive change in the company's internal project to produce artificial intelligence chips.
This change comes after executives decided in 2022 to abandon the first version of the chip, with the company choosing to purchase billions of dollars' worth of graphics processors from NVIDIA.
Nvidia's GPUs have a near monopoly on an AI process called training, which involves feeding large language models with large amounts of data to teach them how to perform tasks.
The new chip performs a process called inference, which uses the model's algorithms to generate responses to user input.
Reuters reported last year that Meta was also developing another chip that could perform training and inference in a similar way to a GPU.
Last year, the company announced details of its first generation training and reasoning software. The ad describes this version of the slide show as an educational opportunity.