NVIDIAGB200 NVL72
Driving the Era of Single-Rack Exascale AI Supercomputing
Connect 36 Grace CPUs and 72 Blackwell GPUs via rack-scale liquid cooling. With 72 GPU NVLink domains operating as a single large GPU, deliver 30x real-time inference performance and 4x training speed for trillion-parameter LLMs—the scalable foundation for advanced AI data centers.
Unlock real-timeTrillion-parameter model
GB200 NVL72 delivers 130 TB/s low-latency GPU communication via the NVLink Switch System, transforming a single rack into an exascale AI and HPC computer. GB200 Grace Blackwell superchips connect two Blackwell Tensor Core GPUs to one Grace CPU using NVLink-C2C, providing a unified compute platform for AI inference, training, and data processing.
Data source comparison: NVIDIA official benchmarks for GB200 NVL72 vs. HGX H100 (LLM inference and energy efficiency: TTL=50ms real-time scenario, 32,768 input/1,024 output; Data processing: database connection and aggregation workloads based on TPC-H Q4). Predicted performance may vary; refer to the latest official data.
TechnologyBreakthrough
From real-time inference to large-scale training, from liquid-cooled efficiency to data processing, the GB200 NVL72 delivers complete capabilities for next-generation AI and accelerated computing.
Real-time trillion-parameter LLM inference
The second-generation Transformer Engine introduces FP4 precision and, paired with fifth-generation NVLink, delivers 30x real-time inference performance for trillion-parameter language models. The 72-GPU NVLink domain combined with liquid cooling design enables the entire rack to function as a single large GPU, breaking through communication bottlenecks.
Accelerated Large-Scale Training
The faster second-generation Transformer Engine delivers FP8 precision, accelerating large-scale LLM training by up to 4x. The fifth-generation NVLink provides per-GPU bandwidth of 1.8 TB/s. Combined with InfiniBand networking and Magnum IO software, it is built for training massive models like MoE.
Liquid-cooled high-efficiency infrastructure
Liquid-cooled racks increase compute density, reduce footprint, and lower data center energy consumption and carbon footprint. With the same power consumption, they deliver 25 times the performance of H00 air-cooled infrastructure while reducing water usage, enabling high-bandwidth, low-latency GPU communication for large-scale NVLink domains.
Accelerated data processing
With high-bandwidth memory, NVLink-C2C interconnects, and a dedicated decompression engine built into the Blackwell architecture, critical database queries run 18 times faster than on CPU, reducing total cost of ownership (TCO) by approximately 5x to accelerate enterprise analytics workloads.
NVIDIA GB200 NVL72Technical Specifications
Complete parameter overview. Contact our technical consultant for selection advice.
| Settings | 36 Grace CPUs | 72 Blackwell GPUs |
| NVFP4 Tensor Core | 1,440 PFLOPS (sparse) | 720 PFLOPS (dense)* |
| FP8/FP6 Tensor Core | 720 PFLOPS (sparse) ** |
| INT8 Tensor Core | 720 POPS (Sparse) ** |
| FP16/BF16 Tensor Core | 360 PFLOPS (sparse) ** |
| TF32 Tensor Core | 180 PFLOPS (sparse) ** |
| FP32 | 5,760 TFLOPS |
| FP64 / FP64 Tensor Core | 2,880 TFLOPS |
| GPU VRAM | Bandwidth | 13.4 TB HBM3E | 576 TB/s |
| NVLink Bandwidth | 130 TB/s (5th Gen NVLink, 72 GPU NVLink Domain) |
| CPU Cores | 2,592 Arm Neoverse V2 cores |
| CPU & Memory | Bandwidth | 17 TB LPDDR5X | 14 TB/s |
| Superchip | GB200 Grace Blackwell Superchip: 1 Grace CPU + 2 Blackwell Tensor Core GPU High-speed interconnect via NVLink-C2C |
| Thermal Performance and Form | Rack-level liquid cooling design |
| Network | Supports NVIDIA Quantum-X800 InfiniBand and Spectrum-X800 Ethernet NVIDIA BlueField-3 DPU enables efficient scaling across hundreds to thousands of GPUs. |
| Manage software | NVIDIA Mission Control |
* Specifications are displayed in sparse and dense modes. ** represents the sparse mode; dense mode has half the specifications of sparse mode. Specification parameters are sourced from the official NVIDIA product page; actual configurations are subject to factory settings.
Product BrochureDownload
Below is a preview of the NVIDIA GB200 NVL72 product brochure. Click the button to generate and download the PDF version for internal review and selection archival.
GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs via rack-scale liquid cooling. The 72-GPU NVLink domain operates as a single large GPU, delivering 130 TB/s low-latency interconnect and 13.4TB HBM3E memory. Trillion-parameter LLM inference is 30x faster than H100, with training speed up 4x and energy efficiency improved 25x.
| Settings | 36x Grace CPU + 72x Blackwell GPU |
| NVFP4 Tensor Core | 1,440 PFLOPS (sparse) | 720 PFLOPS (dense) |
| FP8/FP6 Tensor Core | 720 PFLOPS (sparse) |
| FP16/BF16 Tensor Core | 360 PFLOPS (sparse) |
| GPU VRAM | Bandwidth | 13.4 TB HBM3E | 576 TB/s |
| NVLink Bandwidth | 130 TB/s (5th Gen) |
| Number of CPU cores | 2,592 Arm Neoverse V2 cores |
| CPU Memory | Bandwidth | 17 TB LPDDR5X | 14 TB/s |
| Superchip | 1x Grace CPU + 2x Blackwell GPU (NVLink-C2C) |
| Thermal Design | Rack-level liquid cooling design |
Brochures are generated in real-time from page content. If your browser blocks the download, allow pop-ups and try again.
Ready to deploy GB200 NVL72?
Huaying Technology provides comprehensive professional services for GB200 NVL72 selection consulting, liquid-cooled data center planning, and end-to-end deployment.
Purchase Intent
Submit your requirements and contact details. A sales consultant will confirm pricing, stock availability, and delivery options with you. Product prices are subject to quotation.