NVIDIA has announced a new class of DGX AI Supercomputers named DGX GH200 at Computex 2023. The new supercomputer is powered by NVIDIA GH200 Grace Hopper Superchips and the NVIDIA NVLink Switch System to create a next-generation model for generative AI apps.
With massive shared memory space uses NVLink interconnect technology with the NVLink Switch System which can combine 256 GH200 Superchips, allowing it to perform as a single GPU system. This technology provides the supercomputer with 1 exaflop of performance and 144 terabytes of shared memory. The new computer will have nearly 500x more memory than the previous generations of NVIDIA supercomputers.
Jensen Huang, founder and CEO of NVIDIA, said with the announcement, “Generative AI, large language models, and recommender systems are the digital engines of the modern economy, DGX GH200 AI supercomputers integrate NVIDIA’s most advanced accelerated computing and networking technologies to expand the frontier of AI.”
The GH200 Superchips revolutionize the conventional CPU-to-GPU PCIe connection by integrating an Arm-based NVIDIA Grace CPU with an NVIDIA H100 Tensor Core GPU within a single unit. This breakthrough is made possible through the implementation of NVIDIA NVLink-C2C chip interconnects. As a result, the bandwidth between the GPU and CPU experiences a remarkable 7x boost when compared to the most advanced PCIe technology available.
This innovative design significantly reduces interconnect power consumption by over 5 times. The GH200 Superchips also introduce the impressive 600GB Hopper architecture GPU building block, tailored specifically for DGX GH200 supercomputers.
DGX GH200 is the first of its kind to pair Grace Hopper Superchips with the NVIDIA NVLink Switch System. The previous generation of computers only provided eight GPUs to be combined with NVLink as one GPU without compromising performance.
The new supercomputer provides .48 times more NVLink bandwidth than the previous models, allowing it to become a massive AI supercomputer with the simplicity of programming a single GPU.
Google Cloud, Meta, and Microsoft are some of the first to get access to the DGX GH200 to explore its generative AI workloads. Earlier Microsoft also signed a 10 years long licensing deal with NVIDIA. The company is also expected to provide the design of a new supercomputer as a blueprint for cloud services and other hyperscalers for them to further customize for their infrastructure.
Mark Lohmeyer, vice president of Compute at Google Cloud said, “Building advanced generative models requires innovative approaches to AI infrastructure, The new NVLink scale and shared memory of Grace Hopper Superchips address key bottlenecks in large-scale AI and we look forward to exploring its capabilities for Google Cloud and our generative AI initiatives.”
NVIDIA is also bringing Helios Supercomputer that will be powered by DGX200 technology. The machine will be introduced with NVIDIA Quantum-2 InfiniBand networking to supercharge data throughput for training large AI models.
As NVIDIA announces a new class of DGX AI Supercomputers, the company is continuously working on improving AI integration, NVIDIA recently unveiled ACE for Games at Computex 2023 to bring AI into the world of gaming. The Helios and NVIDIA DGX200 Supercomputers are also expected to be available by the end of this year, marking a beginning of a new era in AI and computer technology.
Source: NVIDIA Newsroom
