PNY NVIDIA DGX Vera Rubin NVL72 20.7TB HBM4 48U Top Feed AI Supercomputer w/ 3 Year Std. Support

NVIDIA DGX Vera Rubin NVL72 Top Feed

The NVIDIA DGX Vera Rubin NVL72 Top Feed is a turnkey, rack-scale AI infrastructure solution purpose-built for gigascale AI training and inference. Built on the NVIDIA Vera Rubin platform, it unifies 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs through sixth-generation NVIDIA NVLink, creating a massive shared-memory fabric for demanding AI workloads. With 20.7TB of HBM4 GPU memory, 1,400TB/s of total GPU memory bandwidth and a fully liquid-cooled architecture, the NVL72 is designed to accelerate large-scale AI workloads while addressing the power and efficiency requirements of modern data centres.

72 Rubin GPUs for Gigascale AI Performance

The DGX Vera Rubin NVL72 combines 72 NVIDIA Rubin GPUs with 36 NVIDIA Vera CPUs to provide a high-density platform for large language models, mixture-of-experts models, video generation, graph neural networks, physical AI simulations and other demanding AI pipelines. The system delivers up to 3,600 PFLOPS of NVFP4 inference performance, 2,520 PFLOPS of NVFP4 training performance and 1,260 PFLOPS of FP8/FP6 training performance, providing substantial accelerated compute for large-scale AI workloads.

20.7TB of HBM4 GPU Memory

With 20.7TB of HBM4 GPU memory, the NVL72 provides a large shared memory resource for processing complex AI models, reasoning workloads and simulations. NVIDIA specifies total GPU memory bandwidth of up to 1,400TB/s in the DGX Vera Rubin NVL72 datasheet, helping deliver the high memory throughput required for large-scale training and inference.

Sixth-Generation NVLink Unifies GPU Compute

Sixth-generation NVIDIA NVLink connects all 72 Rubin GPUs within a single NVLink domain, enabling high-speed GPU-to-GPU communication across the rack. Nine NVIDIA L1 NVLink Switches provide a 216TB/s NVLink bandwidth fabric, allowing the system to operate as a unified accelerated computing platform for workloads that require extensive communication between GPUs.

800Gb/s Networking for Scale-Out AI

The NVL72 provides high-speed networking for integration into larger AI infrastructure. The system incorporates 144 single-port NVIDIA ConnectX-9 VPI interfaces with 800Gb/s NVIDIA InfiniBand and Ethernet, alongside 18 dual-port NVIDIA BlueField-4 VPI interfaces supporting 400Gb/s InfiniBand and Ethernet. This networking architecture provides the scale-out connectivity required for multi-system AI deployments and NVIDIA DGX SuperPOD environments.

Fully Liquid-Cooled Rack-Scale Architecture

Designed as a 48U rack-scale system, the DGX Vera Rubin NVL72 uses a 100% liquid-cooled architecture to support its high-density compute configuration. This approach is designed for data centre environments where power and thermal efficiency are important considerations, while providing a pre-integrated infrastructure platform rather than requiring customers to assemble the individual hardware and software components themselves.

Integrated AI Software and Management

NVIDIA Mission Control provides tools for configuring the infrastructure, managing clusters and workloads, monitoring system health and coordinating operational events across the AI factory. The integrated software stack also includes NVIDIA AI Enterprise and NVIDIA DGX OS, while the host baseboard management controller provides dedicated management connectivity through RJ-45. The platform also supports Ubuntu.

E1.S Storage for AI Infrastructure

The DGX Vera Rubin NVL72 incorporates E1.S storage within its rack-scale architecture, providing a compact enterprise storage form factor suited to high-density data centre infrastructure. Combined with the system's large GPU memory pool and high-speed networking, this supports the data requirements of large-scale AI training and inference deployments.

Three Years of Enterprise Support

The NVIDIA DGX Vera Rubin NVL72 Top Feed comes with three years of enterprise business-standard support for hardware and software, providing manufacturer-backed assistance throughout the standard support period.

Built for Enterprise AI Factories

The NVL72 combines compute, HBM4 memory, NVLink, high-speed networking, liquid cooling and an integrated NVIDIA software stack into a unified rack-scale platform. It is designed to provide a validated foundation for enterprise AI factories and can scale with NVIDIA DGX SuperPOD as organisations expand AI infrastructure across multiple racks.

• Rack-scale 48U AI infrastructure platform.
• Top Feed configuration.
• 72 x NVIDIA Rubin GPUs.
• 36 x NVIDIA Vera CPUs.
• 20.7TB HBM4 GPU memory.
• Up to 1,400TB/s GPU memory bandwidth.
• 3,600 PFLOPS NVFP4 inference performance.
• 2,520 PFLOPS NVFP4 training performance.
• 1,260 PFLOPS FP8/FP6 training performance.
• 9 x NVIDIA L1 NVLink Switches.
• 216TB/s NVIDIA NVLink bandwidth.
• 144 x ConnectX-9 VPI 800Gb/s interfaces.
• 18 x dual-port BlueField-4 VPI interfaces.
• 800Gb/s InfiniBand and Ethernet.
• 400Gb/s InfiniBand and Ethernet.
• 100% liquid-cooled architecture.
• E1.S storage.
• RJ-45 host BMC management connection.
• NVIDIA Mission Control.
• NVIDIA AI Enterprise.
• NVIDIA DGX OS.
• Ubuntu support.
• Three-year enterprise business-standard hardware and software support.