NVIDIA DGX GB200 NVL72 ( AI Supercomputer with 72 Blackwell GPUs and 36 Grace CPUs)

Description

The NVIDIA DGX GB200 NVL72 is a revolutionary rack-scale AI supercomputer designed to power the next era of generative AI, accelerated computing, and large-scale data analytics. Built on the groundbreaking NVIDIA Blackwell architecture, it combines 72 Blackwell GPUs and 36 Grace CPUs into a single, liquid-cooled NVLink-connected system that acts as one massive GPU—delivering unprecedented performance, memory bandwidth, and energy efficiency.

This system is purpose-built for real-time inference of trillion-parameter large language models (LLMs), massive-scale training, and high-throughput data processing. With innovations like the second-generation Transformer Engine, FP4 and FP8 precision, and NVLink-C2C interconnects, the DGX GB200 NVL72 redefines what’s possible in AI infrastructure.

Key Capabilities

Hardware Architecture

• GPU Configuration: 72× NVIDIA Blackwell GPUs
• CPU Configuration: 36× NVIDIA Grace CPUs (2,592 Arm Neoverse V2 cores)
• Memory:
  • Up to 13.5TB HBM3E
  • Up to 17TB LPDDR5X with ECC
  • Up to 30.5TB fast-access memory
• GPU Interconnect:
  • NVLink domain bandwidth: 130TB/s
  • NVLink-C2C: 900GB/s bidirectional bandwidth between Grace and Blackwell
  • 5th Gen NVLink: 1.8TB/s GPU-to-GPU interconnect
• Cooling: Liquid-cooled rack for high-density, energy-efficient operation

Performance Highlights

The DGX GB200 NVL72 is optimized for NVIDIA’s full AI software ecosystem, enabling seamless development, deployment, and scaling of generative AI workloads:

• NVIDIA AI Enterprise – End-to-end software platform for model development, deployment, and monitoring
• NVIDIA NIM™ Microservices – Fast, secure, and scalable model inference
• NVIDIA Magnum IO™ – High-performance data movement and IO stack
• TensorRT™-LLM & NeMo Framework – Accelerated inference and training for LLMs and MoE models
• Confidential Computing – Hardware-based security for sensitive data and models

Use Case Scenarios

• Real-time inference of trillion-parameter LLMs
• Training massive mixture-of-experts (MoE) models
• Enterprise-scale generative AI deployment
• High-throughput database analytics and decompression
• Sustainable AI infrastructure with reduced carbon footprint

Sustainability & Efficiency

The DGX GB200 NVL72 is engineered for sustainability. Its liquid-cooled design reduces water usage and floor space while delivering 25× the performance of air-cooled H100 systems at the same power envelope. This enables enterprises to scale AI workloads without compromising on energy efficiency or environmental impact.

Technical Summary

Last update at December 2025