NVIDIA has unveiled its next-generation AI supercomputer, the NVIDIA DGX SuperPOD, specifically designed to tackle the immense computational requirements of generative AI.
This powerhouse system leverages the processing prowess of NVIDIA GB200 Grace Blackwell Superchips. Each DGX GB200 system houses 36 of these Superchips, which integrate 36 NVIDIA Grace CPUs and a staggering 72 NVIDIA Blackwell GPUs.
Thanks to fifth-generation NVIDIA NVLink technology, these Superchips function cohesively as a single, high-performance supercomputer.
The heart of the DGX SuperPOD
The Grace Blackwell-powered DGX SuperPOD offers immense scalability. It starts with at least eight DGX GB200 systems and easily scales to incorporate tens of thousands of GB200 Superchips interconnected via NVIDIA Quantum InfiniBand.
Customers seeking to power next-generation AI models with massive shared memory requirements can configure an eight DGX GB200 system cluster, linking 576 Blackwell GPUs via NVLink.
- Each DGX GB200 system boasts 36 NVIDIA GB200 Grace Blackwell Superchips
- These superchips are connected to function as a single supercomputer
- GB200 Superchips deliver an impressive 30x performance boost compared to the NVIDIA H100 Tensor Core GPU
Get to know GB200
The GB200 Superchip, featuring a potent combination of Grace CPUs and Blackwell GPUs, is specifically architected to meet the demanding computational needs of large-scale generative AI models. In these complex models, the close integration of both CPU and GPU elements is crucial for optimal performance.
Generative AI models are rapidly growing in size. GB200 is tailored to handle the intricate training and inference workloads for models that operate at the trillion-parameter scale and beyond.
GB200 Superchips deliver an astonishing 30x performance increase compared to the NVIDIA H100 Tensor Core GPU in large language model inference. This kind of performance leap is essential to keep pace with the rapid evolution of generative AI.
The rack-scale architecture of the DGX SuperPOD is optimized for maximum performance by leveraging the GB200 Superchip’s capabilities. This ensures that generative AI workflows are seamlessly handled at scale.
Scaling to meet unprecedented demands
NVIDIA has introduced a new, liquid-cooled, rack-scale architecture optimized for the demands of generative AI. Alongside NVIDIA BlueField-3 DPUs and NVIDIA Quantum-X800 InfiniBand networking, this creates a unified compute fabric.
The platform delivers a staggering 1,800 gigabytes per second of bandwidth to each GPU. Furthermore, fourth-generation NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARP) provides 14.4 teraflops of in-network computing – a fourfold performance boost over the previous architecture.
Complete AI supercomputing solution
The DGX SuperPOD delivers a comprehensive AI supercomputing solution.
To handle the complexities of generative AI workloads, it seamlessly integrates with high-performance storage solutions from NVIDIA-certified partners. To streamline deployment, each system is assembled, wired, and tested at the factory.
Intelligent predictive-management capabilities continuously monitor thousands of hardware and software data points, maximizing uptime and minimizing downtime.
Availability and beyond
NVIDIA also announced the NVIDIA DGX B200 system, a foundational unified AI supercomputing platform utilized globally across industries.
NVIDIA’s DGX SuperPOD with GB200 Superchips promises to accelerate advances in generative AI. From drug development to scientific simulations and content creation, the potential applications of this technology promise to reshape industries.
NVIDIA’s AI journey
NVIDIA’s DGX SuperPOD with GB200 Superchips represents a significant milestone in the company’s ongoing AI journey.
NVIDIA has consistently tested the limits of what’s possible in supercomputing, fueling innovation across a broad spectrum of industries. The capabilities unleashed by this latest system pave the way for even more transformative breakthroughs in the exciting field of generative AI.
NVIDIA DGX SuperPOD with DGX GB200 and DGX B200 systems are expected to be available later this year from NVIDIA’s global partners.
Featured image credit: Kerem Gülen/DALL-E 3.
