NVIDIA GB200 NVL72
The NVIDIA GB200 NVL72 is a rack-scale, liquid-cooled AI supercomputer that integrates 36 Grace CPUs and 72 Blackwell GPUs, interconnected via fifth-generation NVLink, delivering 130 TB/s of GPU communication bandwidth. It provides up to 1,440 PFLOPS of FP4 AI performance and supports up to 13.5 TB of HBM3e GPU memory with 576 TB/s bandwidth.
Equipment NVIDIA — These are highly productive data processing platforms, developed in accordance with modern requirements for computing systems in the fields of artificial intelligence, research, industrial automation and corporate analytics. Architecture used — Hopper, Grace Hopper, Blackwell — provide high calculation density, energy efficiency and scalability.
NVIDIA solutions are focused on comprehensive support for neural network training and inference tasks, including large language models, generative artificial intelligence, computer vision, natural language processing, modeling and virtualization of processes.
Key features of platforms:
-
Support for Hopper, Grace, Blackwell and other architectures Orin
-
The use of graphic processors with interfaces NVLink/NVSwitch for an accelerated interaction between GPU
-
Memory HBM3 or HBM3e with high bandwidth
-
Ability to build scaled clusters, combined into a single computing knot
-
Compatibility with NVIDIA AI Enterprise, CUDA, Triton Inference Server, TensorRT, RAPIDS and other libraries
Scope of equipment NVIDIA:
-
Training and Infession of Modeling and Artificial Intelligence Models
-
Modeling and simulating processes in scientific and engineering tasks
-
Automation of production processes and robotic complexes
-
Creating digital doubles of objects and technological chains
-
Treatment of large data arrays and highly loaded analytical calculations
-
Graphic rendering, 3D modeling, physical process simulation and visualization
A brief overview of the main lines of equipment:
-
HGX — Server platforms for data centers designed to build scalable artificial intelligence infrastructure. Used in rack solutions and dated centers.
-
DGX — Ready NVIDIA Systems for AI-Clea. These are pre -configured solutions with maximum productivity to teach large models. Applied in scientific centers, laboratories, corporate sector.
-
IGX Orin — Industrial-grade platform for embedded artificial intelligence systems. Used in medicine, automation, transport, safety systems.
-
GH200 / GB200 / GB300 — new generation superflies that combine CPU and GPU in a single computing module with coherent memory. Applied in AI cluster, cloud solutions, LLM-models and digital doubles.
-
RTX Workstation — Professional graphics processors for workstations. Designed for design and graphics professionals, 3D, CAD and visualization. Provide accelerated processing of professional applications and AI-tasks.
-
GeForce RTX for a note TbEven — mobile graphic processors designed for resource -intensive tasks: Games, Rendering, Modeling, Application AI at User level.
Advantages of use of equipment NVIDIA:
-
Maximum productivity for artificial intelligence and machine learning tasks
-
Infrastructure optimization to work with new generation models
-
High scalability — from one node to II cluster
-
Compatibility with leading frameworks and software
-
Support for corporate solutions and IT infrastructures
- Perspective for building modern date centers and peripheral computing
| Component | GB200 NVL72 |
|---|---|
| Configuration | 36 Grace CPUs : 72 Blackwell GPUs |
| FP4 Tensor Core | 1,440 PFLOPS (with sparsity) |
| FP8/FP6 Tensor Core | 720 PFLOPS (with sparsity) |
| INT8 Tensor Core | 720 POPS (with sparsity) |
| FP16/BF16 Tensor Core | 360 PFLOPS (with sparsity) |
| TF32 Tensor Core | 180 PFLOPS |
| FP32 | 6,480 TFLOPS |
| FP64 | 3,240 TFLOPS |
| FP64 Tensor Core | 3,240 TFLOPS |
| GPU Memory | Up to 13.5 TB HBM3e |
| GPU Memory Bandwidth | 576 TB/s |
| NVLink Bandwidth | 130 TB/s |
| CPU Core Count | 2,592 Arm® Neoverse V2 cores |
| CPU Memory | Up to 17 TB LPDDR5X |
| CPU Memory Bandwidth | Up to 18.4 TB/s |
| condition | new |
|---|
Log In