AI Research SuperCluster (RSC)
Earlier, Meta announced that it had completed the first phase of its new AI supercomputer RSC (AI Research SuperCluster). When it’s fully assembled later this year, Meta believes it will be the world’s fastest AI supercomputer with “near 5 exaflops of mixed-precision computing power.” In the future, RSC will help develop better AI models that learn from trillions of examples. In Meta’s view, these models will help them build better AR tools and “seamlessly analyze text, images, and videos simultaneously.” Overall, the main value of RSC is to better realize Meta’s vision of the metaverse, in which AI-driven apps and products will play a key role in Meta’s project.
“We hope RSC will help us build new AI systems that, for example, can provide instant speech translation for a large group of people at the same time, so that people who speak different languages can write seamlessly in a research project, or Play the same AR game together,” Meta technical manager Kevin Lee and software engineer Shubho Sengupta wrote on their official blog.
At this stage, RSC has a total of 760 sets of NVIDIA DGX A100 systems as computing nodes, with a total of 6,080 GPUs. Meta said that this version is actually one of the fastest AI supercomputers in the world. According to the initial benchmark results, RSC can perform machine vision tasks up to 20 times faster than Meta’s previous solution, running NVIDIA Collective. Communication Library is also more than 9 times faster than before, and 3 times faster when training large-scale natural language operations.
On this basis, AI models that determine “whether an action, sound, or image is malignant or benign” can be trained faster. Meta believes that such technology can better protect users on the Facebook, Instagram network, and people in the metaverse. In addition to building RSC’s physical infrastructure and systems, Meta will also ensure that it has the relevant security and privacy controls in place to protect massive amounts of real-world training data. By producing real-world data from systems rather than publicly available datasets, RSC is expected to more effectively put its research into practical applications (such as identifying harmful content).
Meta plans to increase the number of GPUs in RSC to 16,000 this year, which it says will increase the system’s AI training performance by more than 2.5 times. In fact, RSC’s related plans can be traced back to early 2020, and Meta hopes that RSC will eventually be able to train AI models at the exabyte level. “We expect this step-by-step functional change in computing power to not only create more accurate AI models for existing services, but also enable entirely new user experiences, especially in the metaverse,” Lee and Sengupta wrote.
