GTC 2024: Nvidia Reveals Digital Blueprint for Next-Gen Data Centers

Jace Dela Cruz, Tech Times 19 March 2024, 03:03 am

Nvidia's recent unveiling at GTC 2024 shed light on its latest endeavor: a comprehensive digital blueprint for the next generation of data centers, showcasing the intricate process involved in designing and simulating these complex systems.

Nvidia Holds Its GTC: Artificial Intelligence Conference — SAN JOSE, CALIFORNIA - MARCH 18: Nvidia CEO Jensen Huang delivers a keynote address during the Nvidia GTC Artificial Intelligence Conference at SAP Center on March 18, 2024, in San Jose, California. The developer conference is expected to highlight new chips, software, and AI processor technology. (Photo: Justin Sullivan/Getty Images)

Nvidia's Next-Gen Data Centers

Crafting modern data centers involves a myriad of factors, from performance and energy efficiency to scalability, necessitating the collaboration of highly skilled engineers versed in various domains, such as compute and network design, computer-aided design (CAD) modeling, and mechanical, electrical, and thermal design.

At the heart of this unveiling is a large cluster built upon the NVIDIA GB200 NVL72 liquid-cooled system. It comprises two racks, each housing 18 NVIDIA Grace CPUs and 36 NVIDIA Blackwell GPUs interconnected by fourth-generation NVIDIA NVLink switches.

During the event, Nvidia showcased a fully operational data center in the digital realm through Nvidia Omniverse, a platform facilitating the connection and development of generative AI-enabled 3D pipelines, tools, applications, and services.

Digital Twin Methodology

Nvidia leveraged its digital twin methodology, constructing a virtual replica of the data center using Omniverse-connected software tools to accelerate the deployment of new data centers.

Engineers could combine and visualize multiple CAD datasets with precision and photorealism through this approach, leveraging the Cadence Reality digital twin platform powered by Nvidia Omniverse APIs.

Replacing an existing cluster within Nvidia's legacy data centers, the GB200 cluster's digital counterpart was crafted using highly accurate point cloud data and panorama photos obtained by scanning the facility with the NavVis VLX wearable lidar scanner by Kinetic Vision.

Subsequent steps involved utilizing Prevu3D software to transform the point cloud into a 3D mesh, providing a detailed digital model of the facility for simulation purposes.

Connected to Omniverse

According to Nvidia, the Cadence Reality platform, seamlessly connected to Omniverse, made the integration of multiple CAD datasets with enhanced realism possible. This platform empowers engineering teams to develop OpenUSD-based 3D tools, workflows, and applications.

Furthermore, the Omniverse Cloud APIs facilitated interoperability with additional tools such as PATCH MANAGER and NVIDIA Air, enabling tasks like designing the physical layout of the cluster and networking infrastructure and optimizing cabling lengths and routing configurations.

Leveraging Cadence's Reality Digital Twin solvers, accelerated by NVIDIA Modulus APIs and NVIDIA Grace Hopper, engineers leveraged simulated various aspects such as airflows and the performance of liquid-cooling systems from partners like Vertiv and Schneider Electric.

Through simulations, the team could optimize cooling systems integrated into the GB200 trays using solutions from Ansys, seamlessly integrating simulation data into the digital twin.

The demonstration underscored the utility of digital twins in thoroughly testing, optimizing, and validating data center designs before physical implementation, as per Nvidia's report.

By visualizing data center performance within the digital twin, engineering teams can streamline design optimization and prepare for diverse scenarios.