NVIDIA launches DGX GH200 AI supercomputer

New AI supercomputer connects 256 Grace Hopper superchips into a massive, 1-Exaflop, 144TB GPU for giant models that power generative AI, recommendation systems and data processing

COMPUTEX - May 29, 2023 - NVIDIA today announced the launch of a new large-memory AI supercomputer - the NVIDIA DGX™ supercomputer powered by the NVIDIA® GH200 Grace Hopper superchip and the NVIDIA NVLink® Switch System, designed to help Develop massive, next-generation models for generative AI language applications, recommendation systems, and data analysis workloads.

The large shared memory space of NVIDIA DGX GH200 connects 256 GH200 super chips through NVLink interconnect technology and NVLink Switch System, allowing them to operate as a single GPU. It offers 1 exaflop of performance and 144 TB of shared memory - nearly 500 times larger than the memory of the previous generation NVIDIA DGX A100 launched in 2020.

NVIDIA founder and CEO Jensen Huang said: "Generative AI, large-scale language models and recommendation systems are the digital engines of the modern economy. The DGX GH200 AI supercomputer integrates NVIDIA's most advanced accelerated computing and network technologies to expand the frontier of AI. "

NVIDIA NVLink technology scales AI at scale

The GH200 superchip uses the NVIDIA NVLink-C2C chip interconnect to integrate the Arm-based NVIDIA Grace™ CPU with the NVIDIA H100 Tensor Core GPU, eliminating the need for traditional CPU-to-GPU PCIe connections. This increases bandwidth between the GPU and CPU by 7x, reduces interconnect power consumption by more than 5x compared to the latest PCIe technology, and provides a 600GB Hopper architecture GPU building block for the DGX GH200 supercomputer.

DGX GH200 is the first supercomputer to pair the Grace Hopper super chip with the NVIDIA NVLink Switch System, a new interconnect that allows all GPUs in the DGX GH200 system to operate together as a whole. The previous generation system could only integrate 8 GPUs into one GPU through NVLink without affecting performance.

The DGX GH200 architecture increases NVLink bandwidth by more than 48 times compared to the previous generation, enabling large-scale AI supercomputer capabilities to be provided through simple programming on a single GPU.

New research tools for AI pioneers

Google Cloud, Meta and Microsoft are among the first companies expected to plug into the DGX GH200 to explore its capabilities for generative AI workloads. NVIDIA also intends to make the DGX GH200 design available as a blueprint to cloud service providers and other hyperscalers so that they can further customize it to their own infrastructure.

"Building advanced generative models requires innovative AI infrastructure," said Mark Lohmeyer, Vice President of Cloud Computing at Google. "The Grace Hopper superchip's new NVLink and shared memory solve key bottlenecks for large-scale AI, and we look forward to its implementation in Google Cloud and our Develop powerful capabilities in generative AI programs.”

Alexis Björlin, Vice President of Infrastructure, AI Systems and Acceleration Platforms at Meta, said: "As AI models become larger and larger, they require powerful infrastructure that can scale to meet growing demand. NVIDIA's Grace Hopper design looks to be able to Let researchers explore new ways to solve their greatest challenges."

Girish Bablani, corporate vice president of Azure Infrastructure at Microsoft, said, “Training large AI models has been a resource- and time-intensive task in the past. The potential of DGX GH200 to process terabyte-level data sets enables developers to operate at a larger scale and Conduct high-level research at a faster pace.”

New NVIDIA Helios supercomputer will advance research and development

NVIDIA is building its own DGX GH200-based AI supercomputer to support the work of the R&D team.

The supercomputer, called NVIDIA Helios, will be equipped with four DGX GH200 systems. Each will be interconnected via an NVIDIA Quantum-2 InfiniBand network to increase data throughput for training large AI models. Helios will contain 1,024 Grace Hopper superchips and is expected to be online by the end of this year.

Fully integrated and built for giant models

The DGX GH200 supercomputer includes NVIDIA software to provide a turnkey, full-stack solution for the largest AI and data analytics workloads. NVIDIA Base Command™ software provides AI workflow management, enterprise-class cluster management and multiple libraries to accelerate compute, storage and network infrastructure, as well as system software optimized for running AI workloads.

It also includes NVIDIA AI Enterprise, the software layer of the NVIDIA AI platform. It provides more than 100 frameworks, pre-trained models, and development tools to simplify the development and deployment of production AI such as generative AI, computer vision, and speech AI.

Availability

The NVIDIA DGX GH200 supercomputer is expected to be available by the end of this year.

Watch Jen-Hsun Huang introduce the NVIDIA DGX GH200 supercomputer during the COMPUTEX 2023 keynote.