NVIDIA - DGX GB200 NVL72 - Rack-Scale Liquid-Cooled AI Supercomputer

The NVIDIA GB200 NVL72 is a rack-scale, liquid-cooled exascale computer designed for real-time trillion-parameter large language model (LLM) inference and massive-scale AI training. It integrates 36 Grace CPUs and 72 Blackwell GPUs into a single NVLink domain, functioning as a unified, massive GPU for high-performance computing (HPC) workloads. - Architecture: NVIDIA Blackwell with 72 GPUs and 36 Grace CPUs interconnected via NVLink-C2C. - Interconnect: Fifth-generation NVIDIA NVLink providing 130 TB/s aggregate GPU communication bandwidth. - Performance: Delivers up to 1,440 PFLOPS FP4 Tensor Core performance and 720 PFLOPS FP8. - Memory: 13.4 TB HBM3E GPU memory with 576 TB/s aggregate memory bandwidth. - CPU Specs: 2,592 Arm Neoverse V2 cores with 17 TB LPDDR5X system memory and 14 TB/s bandwidth. - Cooling: Integrated liquid-cooling system designed to reduce carbon footprint and increase compute density. - Networking: Support for NVIDIA Quantum-X800 InfiniBand and Spectrum-X800 Ethernet platforms. - Inference: 30x faster real-time LLM inference performance for trillion-parameter language models. - Training: 4x faster training performance for large-scale mixture-of-experts (MoE) architectures. - Management: NVIDIA Mission Control for full-stack intelligence and infrastructure resilience.