META announces AI Super Computer for Metaverse
Link to the original (Posted date: 2022/02/01)
META has announced AI RESEARCH SUPERCLUSTER (RSC) supercomputers to accelerate AI research and support the company's metahabus construction.RSC will work in hundreds of different languages from the company, build new and better AI models, and develop new expanded reality tools.
Developing the next -generation advanced AI requires a strong new computer that can operate hundreds of billions per second.Meta researchers have begun to train a large -scale model for research on natural language processing (NLP) and computer vision research using his RSC.It aims to train models using several trillion parameters in the entire Meta business.It is an expansion reality that can be used someday from the content modification algorithm used to detect hate speech on Facebook and Instagram.RSC can train models using multi -modal signals to determine whether the action, sound, or image is harmful or benign.META argues that this will help you keep people safely, not just today's Meta services.
META also defines computer abilities in a different way from the conventional supercomputer measurement methods, because Meta also depends on the performance of the graphics processing chips.This helps to understand the contents of the image, analyze text, and translate between languages.
AI Super Computers build computing nodes by combining multiple GPUs, connect them with high -performance network fabric, enabling high -speed communication between GPUs.RSC is currently composed of 760 NVIDIA DGX A100 systems for a total of 6,080 GPUs as computing nodes.Each DGX communicates through the NVIDIA Quantum 1600 GB/S INFINIBAND 2 level CLOS fabric without overloading.The RSC storage hierarchy includes 175 -petite PewRe Storage Flashurray, 46 -petite Penguin Computing Altus Cash Storage, and 10 -petite Pure Storage Flashblade.
By the end of 2022, the RSC will include a total of about 16,000 GPUs, and will be able to train AI systems using more than 1 trillion parameters in exercises datasets.The GPU number of this element only represents a narrow indicator of the overall performance of the system.For example, Microsoft's AI Super Computer, built at the research institute Openai, is built from 10,000 GPUs.
Building a next -generation AI infrastructure using RSC will help you enhance metahares and make AI more advanced.
