New Class of AI Supercomputer Connects 256 Grace Hopper Superchips to a Massive 1 Exaflop 144TB GPU to Enable Giant Models to Power Generative AI, Recommender Systems and Data Processing
Computex—NVIDIA today unveiled a new class of large-memory AI supercomputers. NVIDIA DGX™ Equipped with a supercomputer Nvidia® GH200 Grace Hopper Super Chip and the NVIDIA NV Link® switch system — Created to enable the development of huge next-generation models for generative AI language applications, recommender systems, and data analytics workloads.
of NVIDIA DGX GH200The large shared memory space uses the NVLink switch system and NVLink interconnect technology to combine 256 GH200 superchips and enable them to run as a single GPU. This provides 1 exaflops of performance and 144 terabytes of shared memory. That’s almost 500x more memory than his previous-generation NVIDIA DGX A100, which was introduced in 2020.
“Generative AI, large language models, and recommender systems are the digital engines of the modern economy,” said NVIDIA founder and CEO Jensen Huang. “The DGX GH200 AI supercomputer integrates NVIDIA’s leading accelerated computing and networking technologies to expand the frontiers of AI.”
NVIDIA NVLink Technology Scales AI at Scale
Combined with Arm-based PCIe connectivity, the GH200 superchip eliminates the need for traditional CPU-to-GPU PCIe connectivity. NVIDIA Grace™ CPUs and NVIDIA H100 Tensor Core GPU Within the same package, using NVIDIA NVLink-C2C Chip interconnect. This delivers a 7x increase in GPU-to-CPU bandwidth and more than a 5x reduction in interconnect power consumption compared to the latest PCIe technology, powering the DGX GH200 supercomputer with a 600GB hopper architecture GPU building block. Provided.
The DGX GH200 is the first supercomputer to combine the Grace Hopper superchip with the NVIDIA NVLink switch system. This is a new interconnect that allows all the GPUs in a DGX GH200 system to work together as one. In previous generation systems, he was able to combine 8 GPUs as 1 GPU with NVLink without compromising performance.
The DGX GH200 architecture delivers 48x more NVLink bandwidth than the previous generation, delivering the power of massive AI supercomputers with the simplicity of programming a single GPU.
New research tools for AI pioneers
Google Cloud, Meta, and Microsoft are expected to be the first to get access to the DGX GH200 to explore the capabilities of generative AI workloads. NVIDIA also plans to make the DGX GH200 design available as a blueprint to cloud service providers and other hyperscalers for further customization to their infrastructure.
“Building sophisticated generative models requires innovative approaches to AI infrastructure,” said Mark Lohmeyer, VP of Computing at Google Cloud. “Grace Hopper’s new NVLink scale and shared memory on superchip addresses a major bottleneck in large-scale AI, and we look forward to exploring its capabilities on Google Cloud and our generative AI initiatives.” I have.”
Alexis Björlin, Vice President of Infrastructure, AI Systems and Accelerated Platforms at Meta, said: “His Grace Hopper design at NVIDIA is intended to enable researchers to explore new approaches to solving their biggest challenges.”
Girish Bablani, Corporate Vice President of Azure Infrastructure at Microsoft, said: “The potential of the DGX GH200 to handle terabyte-sized datasets will enable developers to conduct advanced research at greater scale and accelerated speed.”
New NVIDIA Helios Supercomputer Drives R&D
NVIDIA is building its own DGX GH200-based AI supercomputer to power the work of researchers and development teams.
Dubbed NVIDIA Helios, the supercomputer will feature four DGX GH200 systems.each connected to each other NVIDIA Quantum-2 InfiniBand Networking to increase data throughput for training AI models at scale. Helios contains 1,024 Grace Hopper Superchips and will be online by the end of the year.
Fully integrated and purpose-built for giant models
The DGX GH200 supercomputer includes NVIDIA software that provides a turnkey, full-stack solution for the largest AI and data analytics workloads. NVIDIA base command™ software provides AI workflow management, enterprise-grade cluster management, libraries to accelerate compute, storage, and network infrastructure, and system software optimized to run AI workloads.
Also included are NVIDIA AI Enterprise, the software layer of the NVIDIA AI platform. It offers over 100 frameworks, pre-trained models, and development tools to streamline the development and deployment of production AI, including generative AI, computer vision, speech AI, and more.
availability
The NVIDIA DGX GH200 supercomputer is expected to be available by the end of the year.
Watch Huang talk about the NVIDIA DGX GH200 supercomputer. Keynote speech at COMPUTEX.
