Since the "beginning" of the metaverse, domestic and foreign companies have entered, and after Facebook changed its name to Meta, it has made many attempts to get closer to the metaverse. Meta's AI research team has developed an AI supercomputer, and Meta says it will be the world's fastest AI supercomputer by mid-2022 when it is fully completed.
Meta's supercomputer, called AI Research SuperCluster (RSC), is the result of nearly two years of work, and the company says hundreds of people have been involved in the project, including researchers from partners Nvidia, Penguin Computing and Pure Storage.
Image courtesy of Meta
To build a metaverse, software, hardware, applications, etc. are indispensable. In November 2021, AMD announced a business partnership with Meta, which will purchase AMD's chips to meet its huge demand for data centers and computing power after its transformation from "social media" to "meta-universe company".
For the metaverse to operate, it is naturally inseparable from powerful computing power. Meta's AI supercomputer is already up and running in the first phase, with a total of 760 NVIDIA DGX A100 systems as compute nodes, containing 6,080 GPUs, ranking fifth among the world's fastest supercomputers.
It runs computer vision workflows up to 20 times faster than Meta's traditional production and research infrastructure. By the end of 2022, the AI supercomputer (RSC) will complete the second phase, which will contain approximately 16,000 GPUs upon completion. Its mixed-precision calculation performance will be close to 5 exaflops.
Speaking of supercomputers, they are computers with high speeds and large capacities, with super processing power. The first thing that comes to mind is the equipment used in research institutions to process data from space, nuclear physics and climate change.
Chinese supercomputer: Sunway Taihu Light, image courtesy of Sina
In fact, this kind of equipment and Meta's AI supercomputer can be called high-performance computers or HPC, both of which are closer to the data center in size and appearance than a single computer, and both rely on a large number of interconnected processors to exchange data at a very fast speed.
However, according to Bob Sorensen, HPC analyst at Hyperion Research, "AI-based HPC lives in a world that is a little different from traditional HPC." Show that the biggest difference between the two is accuracy.
Because machine learning requires less accuracy than the tasks of traditional supercomputers, Meta's AI supercomputers don't necessarily compare directly with supercomputers that are often seen in the news.
Image from: World Wide Web
Both supercomputers and AI supercomputers are computed using so-called floating-point algorithms, and most supercomputer speeds are calculated using so-called 64-bit-per-second floating-point operations, or FLOP. However, because AI computation requires less precision, AI supercomputers typically compute in 32-bit or even 16-bit FLOPs.
Meta mentions that the mixed-precision computation performance of the AI supercomputer will be close to 5 exaflops (one EFLOPS (exaFLOPS) equals one hundred kyo/10 billion (10 to the 18th power) floating-point operations per second).
Image credit: REUTERS
To develop such an AI supercomputer with enormous computing power, Meta says it will be used to train a range of systems in Meta's business that can help Meta researchers build AI models that can work across hundreds of languages to analyze text, images, and videos; and make it easier to moderate content on platforms like Facebook and Instagram.
The AI supercomputer could also help develop augmented reality capabilities available in the company's future AR hardware. Meta hopes it will help researchers develop AI models that think like the human brain, supporting the rich multidimensional experiences in Metaverse.
Image credit: W&V
Although the specific computing power after the full completion is still unknown, it is undeniable that Meta's AI supercomputer has a good data processing power. Experiences ranging from machine learning to hardware development to designing virtual worlds will be the cornerstone of Meta's metacosm.
Click "Watching"
It is the greatest motivation for us