NVIDIA’s New DGX Supercomputer Powers Generative AI

During his keynote address at Computex, NVIDIA CEO Jensen Huang made several significant announcements, providing insights into the company’s forthcoming DGX GH200 supercomputer and its focus on assisting companies in developing generative AI models.

The DGX GH200 supercomputer leverages the new NVLink Switch System, enabling 256 GH200 Grace Hopper superchips to function as a single GPU. Each of these chips features an Arm-based Grace CPU and an H100 Tensor Core GPU. NVIDIA claims that this configuration empowers the DGX GH200 to deliver a staggering 1 exaflop of performance and possess 144 terabytes of shared memory, representing nearly 500 times the memory capacity of a single DGX A100 system.

To put this achievement into perspective, the current ranking of the Top500 supercomputers features Oak Ridge National Laboratory’s Frontier as the only known exascale system, with a Linmark benchmark performance of almost 1.2 exaflops. This places Frontier’s peak performance at over twice that of Japan’s Fugaku, the second-ranked system.

Essentially, NVIDIA claims to have developed a supercomputer capable of rivaling the most powerful known system globally. (Meta is also constructing a supercomputer that it asserts will be the world’s fastest AI supercomputer upon completion.) NVIDIA emphasizes that the architecture of the DGX GH200 offers ten times more bandwidth than its predecessor, enabling “the power of a massive AI supercomputer with the simplicity of programming a single GPU.”

Several prominent companies have already expressed interest in the DGX GH200. Google Cloud, Meta, and Microsoft are among the first organizations slated to gain access to the supercomputer for testing its capabilities with generative AI workloads. NVIDIA anticipates that DGX GH200 supercomputers will be available by the end of 2023.

Additionally, NVIDIA is constructing its own supercomputer named Helios, which combines four DGX GH200 systems. The company expects Helios to be operational by the end of this year.

During his keynote, Huang also highlighted other generative AI advancements, including developments in the gaming sector. NVIDIA Avatar Cloud Engine (ACE) for Games is a service that developers can utilize to create customized AI models for speech, conversation, and animation. ACE for Games enables non-playable characters to possess conversational skills and respond to questions with evolving lifelike personalities.

NVIDIA’s latest announcements underscore the company’s commitment to advancing the field of generative AI and providing powerful tools and infrastructure to drive innovation in various industries. As the DGX GH200 supercomputer and other advancements become available, organizations can expect further breakthroughs in AI model development and performance.