The new supercomputer developed by Meta is a key element in building the future Metavers VR, but will also handle content published on other company platforms, such as Facebook and Instagram. Meta, the owner of Facebook, thus enters the global race to build the largest and most powerful supercomputers.
At least in the segment of artificial intelligence technologies, the AI Research SuperCluster facility will soon become the most powerful supercomputer in the world. Built in about two years using a team of hundreds of specialists and researchers, including recruits from partners Nvidia, Penguin Computing and Pure Storage, the facility dedicated to accelerating neural processing algorithms is already used to train AI models specializing in language processing. natural and computer-vision technologies for research.
“The experiences we build for metavers require enormous computing power, and CSR will enable new AI models that can learn from trillions of examples, understand hundreds of languages, and more,” said Meta CEO Mark Zuckerberg. in a statement to The Wall Street Journal.
Currently, the AI supercomputer developed by Zuckerberg’s company houses 6080 Nvidia graphics processing units, placing it in fifth place among the fastest supercomputers in the world. But in the final form, it will have about 16,000 GPU-like processors found on NVIDIA’s top-of-the-line graphics accelerators, becoming the world’s most powerful AI supercomputer.
Unlike complex simulations and analyzes usually performed using supercomputers, the processing of artificial intelligence algorithms using conventional systems equipped with x86 or ARM processors is rather inefficient. In contrast, GPUs such as those used to accelerate 3D gaming are more suitable for this type of task, based on performing a very large number of relatively simple calculations in parallel.
Compared to a conventional CPU, equipped with only a few very powerful computing units but expensive and inefficient in terms of power consumption, a GPU has thousands of processors that can run simultaneously. Based on a simplified design, they are less versatile in terms of the type of code they can execute, but much more powerful and efficient in the type of task used in this case.
High-performance computing infrastructure, such as a supercomputer, is needed to understand the amount of data Facebook needs, said Bill Gropp, director of the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign.
“In the case of Facebook, I suspect that what they are looking at is the treasure they have from the data they get from their users, and this is a huge data set, too big for most artificial intelligence systems that researchers I usually use them. “Dr. Gropp said.
According to Meta Vice President Jerome Pesenti, the new AI Research SuperCluster is currently being used for research purposes and could take years to see what kind of products the company now intends to develop. Meta’s goal is to process the vast databases collected so far in order to build AI models that can think like a human brain and interact using voice and visual recognition, benefiting from contextual understanding of situations.
For now, Meta refuses to specify the location of the new supercomputer, or its cost.