Product Description
72 NVIDIA Blackwell Ultra GPUs | 36 NVIDIA Grace CPUs | Rack-Scale AI Supercomputer for Large Model Training & Inference
The NVIDIA GB300 NVL72 by HPE is a next-generation rack-scale AI platform engineered to accelerate the training, fine-tuning, and real-time inference of AI models with over 1 trillion parameters. Combining 72 NVIDIA Blackwell Ultra GPUs, 36 NVIDIA Grace CPUs, high-speed NVLink interconnects, advanced InfiniBand networking, and proven HPE Direct Liquid Cooling (DLC) technologies, it delivers unprecedented performance for hyperscale AI, large language models (LLMs), deep learning, and high-performance computing (HPC) workloads.
Built for AI service providers, research organizations, cloud providers, and enterprises developing advanced AI models, the NVIDIA GB300 NVL72 by HPE provides a complete, scalable platform that combines breakthrough computing power, efficient cooling, high-speed communications, integrated software, and enterprise-grade support services.
Key Features
- Optimized for AI models exceeding 1 trillion parameters
- 72 NVIDIA Blackwell Ultra GPUs interconnected via 5th-generation NVIDIA NVLink
- 36 NVIDIA Grace CPUs for accelerated AI and HPC workloads
- Up to 20TB HBM3E GPU memory across the system
- 800G networking powered by NVIDIA ConnectX-8 SuperNIC and XDR InfiniBand
- Direct Liquid Cooling (DLC) technology from HPE
- Up to 90% liquid-cooled rack architecture
- Designed for large-scale AI training, fine-tuning, and real-time inference
- Integrated NVIDIA CUDA, cuDNN, AI frameworks, and management tools
- Accelerated deployment with HPE AI infrastructure expertise
- Comprehensive HPE lifecycle services and support
- Enterprise-ready architecture for hyperscale AI environments
Specifications
- NVIDIA GB300 NVL72 by HPE
- Rack-Scale AI Computing Platform
- 72 x NVIDIA Blackwell Ultra GPUs
- 36 x NVIDIA Grace ARM CPUs
- 5th Generation NVIDIA NVLink Interconnect
- 18 x 1U Compute Trays
- 9 x 1U NVLink Switch Trays
- 8 x 1U 33kW Power Shelves
- 48RU Rack Configuration
- 288GB HBM3E Memory per GPU
- 8TB/s Memory Bandwidth per GPU
- Up to 20TB Total HBM3E Memory
- Up to 569TB/s Aggregate Memory Bandwidth
- 2 x NVIDIA Grace CPUs per Compute Tray
- 4 x NVIDIA Blackwell Ultra GPUs per Compute Tray
- 4 x 128GB SOCAMM per Tray
- 512GB Total Memory per Tray
- 480GB Usable Memory per CPU
- 8 x External E1.S SSDs per Compute Tray
- 1 x Internal M.2 Boot Drive per Compute Tray
- NVIDIA NVLink Switch Fabric
- 1.8TB/s Bidirectional GPU-to-GPU Connectivity
- 18 x 100GB/s Links per GPU
- 800G NVIDIA ConnectX-8 SuperNIC Networking
- NVIDIA XDR InfiniBand Networking
- NVIDIA SHARP Technology
- Direct Liquid Cooling (DLC)
- Hybrid Air and Liquid Cooling Architecture
- Approximately 132kW to 140kW Rack Power Consumption
- 8 x 33kW Power Shelves with N+1 Redundancy
- 1400A 50V DC Busbar Power Distribution
- Cooling Distribution Unit (CDU) Support
- Supports Up to 8 Racks per CDU
- Rack Dimensions: 600mm x 2298mm x 1068mm
- Rack Extensions: Up to 1710mm
- Rack Weight: Approximately 3300lb
AI Software Integration
- NVIDIA CUDA
- NVIDIA cuDNN
- NVIDIA AI Frameworks
- Advanced Cluster Management Tools
- High-Performance Networking Software
- System Health Monitoring
- Leak Detection Management Tools
- Operational Management Interfaces
Ideal For
- Large Language Model (LLM) Training
- Foundation Model Development
- Generative AI Platforms
- AI Service Providers
- Hyperscale AI Infrastructure
- High-Performance Computing (HPC)
- Deep Learning Research
- Real-Time AI Inference
- Enterprise AI Deployments
- Scientific Computing Environments