Meta, the parent company of Facebook, introduced AI Research SuperCluster (RSC), a supercomputer for artificial intelligence (AI) research projects.
Developed 1.5 years ago, RSC’s goal is to create more powerful AI tools. Its purpose is to help with complex tasks like detecting hateful content in Facebook or Instagram posts. Meta says RSC is one of the most powerful AI supercomputers available today. And hopefully it will become the fastest when fully built by the end of the year.
“With RSC, we can train (AI) models using multimodal cues, such as post context recognition based on language, images, and tone of voice,” said Shubho Sengupta, engineer software at Meta said.
With powerful performance, supercomputers are increasingly widely used for AI research. At the US Department of Energy, the IBM Summit supercomputer is applied to analyze the world’s hot issues. Firms like Microsoft and NVIDIA also use supercomputers for their own purposes.
According to The Verge, Meta engineers are responsible for the design of the RSC’s cooling, power, networking, and cabling systems. On the official website, Meta said that as of January, RSC was equipped with 6,080 graphics cards (GPUs) divided into 760 Nvidia DGX A100 nodes (multiple CPUs, GPUs, RAM systems. All were connected to form a supercomputer). Complete).
Currently, the power of RSC is equivalent to the Perlmutter supercomputer located at the US National Research Center for Energy Computer Science , ranked 5th in the world. However, when completed later this year, Meta says the RSC will have 16,000 GPUs with 5 exaflops of compute performance, or 5,000 petaflops (5 billion operations per second), making it the world’s fastest AI supercomputer.
Compared with ordinary supercomputers, AI supercomputers have many similarities such as depending on nodes connecting with each other. The big difference between the two supercomputers lies in the accuracy. Since AI calculations require less precision than tasks on traditional supercomputers, AI supercomputers can perform more calculations per second.
Bob Sorensen, an analyst at Hyperion Research, said that the speeds announced by manufacturers are often the highest performance in theory. In fact, supercomputers are considered efficient when they can get the most out of the tasks they’re designed for.
“Therefore, it is not surprising that some supercomputing systems have less than 25 percent actual performance compared to theoretical performance,” Sorensen said.
Meta’s research team used RSC to train models related to natural language processing and computer vision. In the future, they could use RSC to train models that simultaneously analyze text, images, and video, CNN reported.
Over time, Meta says RSC will help build entirely new AI systems, for complex tasks like real-time translation of large amounts of content in multiple languages, helping people work together. Or play games together even though they speak different languages.
In previous Meta tests, RSC was able to train complex language models 3 times faster than the current system. That means a typical AI model takes 9 weeks to train, the time will be shortened to 3 weeks with RSC.
Meta expects RSC to be able to train AI models with trillions of parameters, in datasets that are EB (exabytes) in size. Ultimately, this supercomputer will play an important role in building the virtual universe (metaverse), the new area the social networking company is aiming for.