NVIDIA unveils next-gen data centre blueprint with advanced AI supercomputers

NVIDIA has revealed its blueprint for constructing the next generation of data centres. The company's advanced AI supercomputers, backed by Omniverse digital twin and supported by Ansys, Cadence, PATCH MANAGER, Schneider Electric and Vertiv, are laying the groundwork for an efficient AI infrastructure. The construction and operation of modern data centres are increasingly elaborate, requiring skilled engineers across multiple disciplines including performance, energy efficiency and scalability considerations.

The most recent addition to NVIDIA's industry-leading AI supercomputers is a substantial cluster built on the NVIDIA GB200 NVL72 liquid-cooled system. The system consists of two racks, each equipped with 18 NVIDIA Grace CPUs and 36 NVIDIA Blackwell GPUs, all interconnected by fourth-generation NVIDIA NVLink switches.

At a recent showcase, NVIDIA demonstrated this fully operational data centre as a digital twin in NVIDIA Omniverse. The platform, designed for building AI-enabled 3D pipelines, tools, applications and services, is where the company first built its digital twin to expedite the launch of new data centres. Engineers collaborated to unify and visualise a range of CAD datasets with unprecedented photorealism using the Cadence Reality digital twin platform, empowered by NVIDIA Omniverse APIs.

NVIDIA's latest GB200 cluster is a replacement for an earlier model in one of the company's legacy data centres. The digital blueprinting process started with tech company Kinetic Vision scanning the facility with the NavVis VLX wearable lidar scanner. This produced high-accuracy point cloud data and panorama images, with Prevu3D software subsequently deployed to convert the point cloud into a 3D mesh – creating a highly accurate model of the facility for simulation purposes.

The engineers successfully harmonised a breadth of CAD datasets with the Cadence Reality platform, offering enhanced precision and realism. Collaborating with Omniverse granted the team a potent computing platform to develop OpenUSD-based 3D tools, workflows and applications. Omniverse Cloud APIs also allowed interoperability with more software, with PATCH MANAGER being used to design the physical layout of their cluster and networking infrastructure, ensuring accurate cabling lengths and routing.

Additionally, the Cadence Reality Digital Twin solvers, accelerated by NVIDIA Modulus APIs and NVIDIA Grace Hopper, were utilised to simulate the performance of the new liquid-cooling systems from Vertiv and Schneider Electric. They integrated cooling systems within the GB200's trays, using Ansys solutions to feed simulation data into the digital twin. The design's performance was then simulated and optimised using this digital twin.

This process illustrated how digital twins enable comprehensive testing, optimisation, and validation of data centre designs before constructing any physical models. Furthermore, by visualising performance data, design teams can better plan and optimise designs. By considering a multitude of variables such as cabling lengths, power, cooling and space in an integrated way, cluster creation can occur faster, with greater efficiency, and with greater optimisation than in the past.