---
格式版本: 2
标题: "Gigascale AI Training & Inference Platform | NVIDIA DGX Vera Rubin NVL72"
原文链接: "https://www.nvidia.com/en-eu/data-center/dgx-vera-rubin-nvl72"
发布日期: "2026-05-22"
发现时间: "2026-05-22T00:00:00+08:00"
入库时间: "2026-05-28T02:39:38+08:00"
来源平台: "历史资料迁移"
搜索渠道: "legacy_migration"
搜索词: "NVIDIA Vera Rubin rack scale AI"
匹配关键词:
  - "AI Rack"
  - "Scale-up"
  - "NVIDIA Vera Rubin rack scale AI"
  - "待定数据"
  - "厂家"
  - "Nvidia"
  - "Gigascale"
  - "AI"
  - "Training"
  - "Inference"
  - "Platform"
  - "NVIDIA"
  - "DGX"
  - "Vera"
  - "Rubin"
  - "NVL72"
  - "22.md"
相关厂家:
  - "NVIDIA"
  - "AWS"
相关专家:
  []
内容类型: "网页"
抓取工具: "Firecrawl 原始抓取 + qwen3.6-flash 原文清洗"
清洗工具: "历史资料原文保留 + LLM 正文裁剪"
关联判断: "待人工确认"
关联置信度: 0
关联理由: "历史资料迁移，未重新调用模型判定"
AI优质: "是"
AI打分: 99
AI分档: "高置信优质"
AI质检状态: "通过"
AI打分理由: "官方一手发布，直接聚焦NVL72机柜级AI超节点架构。提供CES 2026最新Rubin/Vera芯片规格、NVLink拓扑、800G网络及详细算力/内存参数，技术细节与商业部署信号明确，内容完整权威。"
AI质检模型: "qwen3.6-plus"
AI质检时间: "2026-06-03T00:29:14+08:00"
AI主题相关性: 20
AI来源权威性: 15
AI新颖性: 20
AI技术细节: 20
AI商业部署信号: 14
AI完整性: 10
采集批次: "2026年5月28日2点39分33秒"
采集批次ID: "legacy-migration-20260528023933"
去重键: "https://www.nvidia.com/en-eu/data-center/dgx-vera-rubin-nvl72"
---

![NVIDIA DGX Vera Rubin NVL72](./assets/01-www.nvidia.com.jpg)
# NVIDIA DGX Vera Rubin NVL72
The foundation for gigascale AI training and inference.

## Overview
## Scaling Intelligence in the Age of Agentic AI
With the convergence of pretraining, post-training, and inference-time scaling laws, modern data centers face performance and efficiency bottlenecks when trying to support models that require complex, multi-step chains of thought. Purpose-built to meet these challenges, NVIDIA DGX™ Vera Rubin NVL72 is a turnkey AI infrastructure solution that delivers leading-edge performance while vastly improving energy efficiency. DGX Vera Rubin NVL72 delivers industry-leading performance per watt and tokens per watt for any AI workload, allowing enterprises to dramatically scale intelligence while improving token economics.

### NVIDIA DGX SuperPOD Sets the Stage for Rubin-Based Systems
Announced at CES 2026, NVIDIA DGX SuperPOD™ serves as a blueprint for deploying AI factories built on the NVIDIA Rubin architecture.

### Introducing NVIDIA Rubin
Learn more about the NVIDIA Rubin GPU architecture.

### Features
## Accelerated Training and Inference for Any AI Model
### Maximum Efficiency
With NVIDIA Vera Rubin Superchips DGX Vera Rubin NVL72 pairs the power-efficient [NVIDIA Vera CPU](https://www.nvidia.com/en-eu/data-center/vera-cpu/) with [NVIDIA Rubin GPUs](https://www.nvidia.com/en-eu/data-center/technologies/rubin/) to deliver leading performance per watt. This enables enterprises to scale training and long-context inference within their existing energy footprint, turning the data center into an energy-efficient engine for AI innovation.

### The Universal Engine for AI at Scale
[NVIDIA DGX SuperPOD](https://www.nvidia.com/en-eu/data-center/dgx-superpod/) with DGX Vera Rubin NVL72 systems eliminates the guesswork from deploying large-scale AI infrastructure. With a simple, easy-to-deploy rack architecture, it’s designed to take advantage of the latest advancements in data center engineering.

### Accelerating Operations With Intelligent Software
[NVIDIA Mission Control™](https://www.nvidia.com/en-eu/data-center/mission-control/) accelerates every aspect of operations, from configuring DGX Vera Rubin NVL72 to integrating with facilities to managing clusters and workloads. Enabled by intelligent, integrated software, enterprises gain control over cooling and power events to redefine infrastructure resiliency.

## The Proven Standard for Enterprise AI
Built from the ground up for enterprise AI, the NVIDIA DGX platform, featuring NVIDIA DGX SuperPOD, combines the best of NVIDIA software, infrastructure, and expertise in a modern, unified AI development solution.
![The Proven Standard for Enterprise AI](./assets/02-www.nvidia.com.jpeg)

## Specifications
## NVIDIA DGX Vera Rubin NVL72 Specifications
| | |
| --- | --- |
| GPU and CPU | 72x NVIDIA Rubin GPUs, 36x NVIDIA Vera CPUs |
| Total GPU Memory \| Bandwidth | 20.7 TB \| Up to 1,580 TB/s |
| Total Fast Memory | 75 TB |
| Performance | NVFP4 Inference: 3,600 PFLOPS<br> NVFP4 Training: 2,520 PFLOPS*<br> FP8/FP6 Training: 1,260 PFLOPS* |
| Networking | > 144x OSFP single-port NVIDIA® ConnectX®-9 VPI with 800 Gb/s NVIDIA InfiniBand and Ethernet<br> > 18x dual-port NVIDIA BlueField®-4 VPI with 400 Gb/s NVIDIA InfiniBand and Ethernet |
| NVIDIA NVLink™ Switch System | 9x L1 NVIDIA NVLink Switches |
| Management Network | Host baseboard management controller (BMC) with RJ45 |
| Software | NVIDIA Mission Control, NVIDIA AI Enterprise, NVIDIA DGX OS |
| Enterprise Support | Three years of enterprise business-standard support for hardware and software |
Specifications subject to change. *Dense specification.

## NVIDIA DGX Vera Rubin NVL72 Quick Specs
| | |
| --- | --- |
| GPU and CPU | 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs |
| Total GPU Memory \| Bandwidth | 20.7 TB \| Up to 1,580 TB/s |
| NVIDIA NVLink™ Switch Systems | 9x L1 NVIDIA NVLink Switches |
| Total Fast Memory | 75 TB |
| Performance | NVFP4 Inference: 3,600 PFLOPS<br> NVFP4 Training: 2,520 PFLOPS*<br> FP8/FP6 Training: 1,260 PFLOPS* |
Specifications subject to change. *Dense specification.