Meta, the parent company of Facebook, has shared plans for a new data center architecture for AI workloads, along with the development of its own custom chip for running artificial intelligence models.
“This new data center will be an AI-optimized design, supporting liquid-cooled AI hardware and a high-performance AI network connecting thousands of AI chips together for data center-scale AI training clusters,” said Santosh Janardhan head of infrastructure at Meta, in a blog post, adding that the new data center systems will be faster and more cost-effective to build than earlier facilities.
Its new data center design will be optimized to train AI models, a process that enables them to better their performance as they ingest more data.
In addition to the new data center design, the company said that it was working on developing AI supercomputers that will support training of next-generation AI models, power augmented reality tools, and support real-time translation technology.
“We are executing on an ambitious plan to build the next generation of Meta’s AI infrastructure and today, we’re sharing some details on our progress. This includes our first custom silicon chip for running AI models, a new AI-optimized data center design and the second phase of our 16,000 GPU supercomputer for AI research,” said Janardhan.
Meta’s custom chip for running AI models, called Meta Training and Inference Accelerator (MTIA), is designed to provide greater compute power and efficiency than CPUs on the market today, according to Janardhan.
MTIA is customized for internal workloads such as content understanding, feeds, generative AI, and ad ranking, the company said, adding that the first version of the chip was designed in 2020.
For more information about this news, visit https://about.fb.com.