Product Description
Lenovo NVIDIA GB300 NVL72 AI Platform
72 NVIDIA Blackwell GPUs | NVIDIA Grace CPUs | Rack-Scale Enterprise AI Platform
The Lenovo NVIDIA GB300 NVL72 is a rack-scale AI platform engineered to accelerate large-scale AI training, inference, reasoning, and high-performance computing workloads. Built on the powerful NVIDIA Grace-Blackwell architecture, the system combines NVIDIA Grace CPUs and NVIDIA Blackwell GPUs with fifth-generation NVIDIA NVLink technology to create a unified AI infrastructure designed for enterprise-scale deployments.
Designed for modern AI factories and enterprise AI environments, the platform delivers massive GPU acceleration across a unified 72-GPU domain, making it ideal for large language models (LLMs), generative AI, agentic AI, vector databases, and advanced AI inference workloads. High-speed networking, extensive memory bandwidth, and rack-scale NVLink connectivity provide the performance required for demanding AI and HPC applications.
The Lenovo NVIDIA GB300 NVL72 features an advanced hybrid liquid-air cooling architecture that removes heat directly at the source, enabling exceptional compute density while maintaining predictable performance and energy efficiency. Warm-water operation further improves cooling efficiency and reduces operational costs, making it an ideal platform for next-generation data centers.
Key Features
- NVIDIA Grace-Blackwell architecture optimized for enterprise AI and HPC
- Unified 72-GPU rack-scale NVLink domain for large-scale AI workloads
- 5th-generation NVIDIA NVLink technology for coherent CPU-GPU memory access
- 72 NVIDIA Blackwell GPUs interconnected across the rack
- 2 NVIDIA Grace CPUs and 4 NVIDIA Blackwell GPUs per compute tray
- 18 compute trays within a 48U MGX rack architecture
- 9 NVLink switch trays for rack-scale GPU interconnectivity
- HBM3e GPU memory architecture optimized for AI training and inference
- LPDDR5X CPU memory for high-bandwidth data processing
- Integrated 800Gb/s CX8 networking for high-performance communications
- Support for InfiniBand or Ethernet scale-out clustering
- Hybrid liquid-air cooling with warm-water operation from 18°C to 45°C
- Liquid cooling applied to CPUs, GPUs, voltage regulators, and NVLink modules
- Up to 8 x 33kW power shelves with balanced 48V busbar distribution
- NVIDIA OpenBMC-based systems management
- Designed for LLM training, AI inference, vector databases, and agentic AI workloads
- 3-Year Onsite Next Business Day Warranty
Specifications
- Form Factor: 1U Grace-Blackwell Compute Tray
- 18 Compute Trays in a 48U MGX Rack
- 9 x 1U NVLink Switch Trays
- 2 x NVIDIA Grace CPUs per Compute Tray
- 4 x NVIDIA Blackwell GPUs per Compute Tray
- 72 NVIDIA Blackwell GPUs per Rack
- 5th-Generation NVIDIA NVLink Technology
- HBM3e GPU Memory Architecture
- High-Bandwidth LPDDR5X CPU Memory
- Up to 8 x E1.S NVMe SSDs per Compute Tray
- Integrated 800Gb/s CX8 Networking
- Dual 25GbE LOM Connectivity
- Rack-Scale NVLink Domain Across 72 GPUs
- Support for High-Speed InfiniBand or Ethernet Scale-Out Clusters
- Hybrid Liquid + Air Cooling System
- Supports PG25 or Deionized Water Cooling
- Warm Water Operation: 18°C to 45°C
- Up to 8 x 33kW Power Shelves
- Balanced 48V Busbar Power Distribution
- Advanced Monitoring via Confluent and NVIDIA Tools
- NVIDIA OpenBMC-Based Systems Management
- Supported Operating System: Ubuntu 22.04 (HWE)
- Standard Warranty: 3-Year Onsite Next Business Day Service
- Optional Warranty: 5-Year Service Available
Ideal For
- Large Language Model (LLM) Training
- Generative AI Platforms
- AI Reasoning and Agentic AI
- Enterprise AI Infrastructure
- AI Inference at Scale
- Vector Databases
- Deep Learning Workloads
- High Performance Computing (HPC)
- Scientific Computing and Simulations
- Large-Scale AI Factories