Technology company NVIDIA has added a revolutionary AI supercomputer to its data center solutions portfolio. The NVIDIA DGX GH200, powered by the innovative NVIDIA GH200 Grace Hopper Superchips and the NVIDIA NVLink Switch System, are anticipated to revolutionize AI development by enabling the creation of massive, next-generation models for generative AI language applications, recommender systems, and data analytics workloads.
The visionary founder and CEO of NVIDIA, Jensen Huang, emphasized the significance of this innovation, stating, “Generative AI, large language models, and recommender systems are the digital engines of the modern economy. Expanding the frontier of AI, DGX GH200 AI supercomputers incorporate NVIDIA’s most advanced accelerated computation and networking technologies.”
The DGX GH200 supercomputer features a remarkable shared memory space that makes use of NVLink interconnect technology and the NVLink Switch System. By integrating 256 GH200 superchips into a single GPU, this system would deliver an astounding 1 exaflops of performance and 144 terabytes of shared memory. This significant increase corresponds to nearly 500 times more memory than the 2020-introduced previous-generation NVIDIA DGX A100.
The DGX GH200 superchips eradicate the need for traditional CPU-to-GPU PCIe connections. By combining an Arm-based NVIDIA Grace CPU and an NVIDIA H100 Tensor Core GPU within the same package and employing NVIDIA NVLink-C2C chip interconnects, the bandwidth between the GPU and CPU would be increased by sevenfold in comparison to the most recent PCIe technology. This innovation would also reduce the power consumption of interconnects by a factor of five and provide a 600GB Hopper architecture GPU construction element for the DGX GH200 supercomputers.
Google Cloud, Microsoft Azure, Meta
The DGX GH200 is the first supercomputer to integrate Grace Hopper Superchips with the NVIDIA NVLink Switch System, a revolutionary interconnect that enables all GPUs within a DGX GH200 system to function as a single entity. The DGX GH200 architecture would deliver the power of a massive AI supercomputer while retaining the simplicity of programming a single GPU.
Notably, a number of tech titans are among the first to obtain access to the DGX GH200 and investigate its capabilities for AI-generated workloads. Google Cloud, Meta, and Microsoft are planning to exploit this supercomputer’s potential. In addition, NVIDIA plans to share the DGX GH200 design as a blueprint with cloud service providers and other hyperscalers, allowing them to further tailor the technology to their particular infrastructures.
Mark Lohmeyer, Vice President of Compute at Google Cloud, expressed enthusiasm for the initiative, stating, “Developing advanced generative models requires novel AI infrastructure. Grace Hopper Superchips’ new NVLink scale and shared memory address key constraints in large-scale AI, and we look forward to investigating its capabilities for Google Cloud and our generative AI initiatives.”
Alexis Bjorlin, Vice President of Infrastructure, AI Systems, and Accelerated Platforms at Meta, acknowledged the need for robust infrastructure to support the expansion of AI models, emphasizing that NVIDIA’s Grace Hopper design enables researchers to explore new methods for tackling the toughest problems.
Girish Bablani, Microsoft’s Corporate Vice President of Azure Infrastructure, emphasized the DGX GH200’s ability to work with terabyte-sized datasets, enabling developers to conduct sophisticated research on a larger scale and at faster velocities.
