NVIDIADGX GB300
Rack-Level AI Factory Infrastructure for the Era of AI Inference
Integrates 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs to deliver up to 1440 PFLOPS of FP4 performance, 20TB of GPU memory, and 37TB of total fast memory. The liquid-cooled rack design is compatible with NVIDIA MGX racks and can scale to thousands of nodes when combined with DGX SuperPOD, helping enterprises break through performance bottlenecks in large-scale AI.
Enterprise-gradeAI Factory Infrastructure
NVIDIA DGX GB300 is a high-performance, scalable rack-level AI system designed for enterprises. Built on the NVIDIA Grace Blackwell Ultra superchip and featuring liquid cooling with the NVIDIA MGX rack, it scales elastically to thousands of nodes. It supports the complete AI lifecycle—from development to deployment—enabling organizations to run AI factories at the efficiency of hyperscale cloud providers.
Empower every enterpriseUltra-scale AI capabilities
From superchips to elastic scaling, from operational resilience to data center compatibility, DGX GB300 delivers complete capabilities for large-scale AI deployment.
Grace Blackwell Ultra Superchip
A rack-scale liquid-cooled system powered by 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs, designed for training, post-training, and inference at test time. Delivers up to 1.5x denser FP4 performance and 2x better attention performance compared to DGX GB200.
Elasticly scale to thousands of instances
Leveraging the NVIDIA DGX SuperPOD reference architecture, DGX GB300 scales elastically from a single system to thousands of nodes. Its simple-to-deploy rack design helps enterprises overcome performance bottlenecks in large-scale AI.
Outstanding infrastructure resilience
NVIDIA Mission Control operationalizes AI workflows from workloads to infrastructure and delivers enterprise-grade operations as software, providing full-stack intelligence for Blackwell data centers to ensure stable operation of AI factories.
Designed for modern data centers
The rack-level liquid cooling system uses NVIDIA MGX racks, integrates seamlessly into modern data centers, leverages the latest advancements in data center engineering, and simplifies deployment of large-scale AI infrastructure.
NVIDIA DGX GB300Technical Specifications
Complete parameter overview. Contact our technical consultant for selection advice.
| GPU | 72x NVIDIA Blackwell Ultra GPU, 36x NVIDIA Grace CPU |
| CPU cores | 2,592 Arm Neoverse V2 cores |
| GPU VRAM | Bandwidth | 20 TB | Up to 576 TB/s |
| Total Quick Memory | 37 TB |
| Performance | FP4 Tensor Core: 1440 PFLOPS (sparse) | 1080 PFLOPS (dense)* FP8/FP6 Tensor Core: 720 PFLOPS (non-sparse) ** |
| Network | 72 OSFP single-port NVIDIA ConnectX-8 VPI, supporting 800 Gb/s NVIDIA InfiniBand 18 dual-port NVIDIA BlueField-3 VPIs, supporting 200 Gb/s NVIDIA InfiniBand and Ethernet |
| NVLink Switch System | 9x L1 NVIDIA NVLink Switch |
| Manage Network | Host Baseboard Management Controller (BMC), RJ45 port |
| Software | NVIDIA Mission Control NVIDIA AI Enterprise NVIDIA DGX OS / Ubuntu |
| System Type | Rack-scale liquid cooling system built on NVIDIA MGX racks, scalable to thousands of units |
| Support | 3-Year Business Standard Hardware and Software Support Services |
* Specifications are displayed in Sparse/Dense mode; ** indicates no sparse mode. Specifications sourced from the official NVIDIA product page; actual configuration is subject to factory settings.
Product BrochureDownload
Preview of the NVIDIA DGX GB300 product brochure. Click the button to generate and download a PDF version for internal review and selection archiving.
The DGX GB300 is built on the NVIDIA Grace Blackwell Ultra superchip, integrating 72 Blackwell Ultra GPUs and 36 Grace CPUs to deliver 1440 PFLOPS FP4 performance (sparse), 20 TB of GPU memory, and 37 TB of total fast memory. It offers 1.5× higher dense FP4 performance and 2× improved attention performance compared to the DGX GB200, with scalability to thousands of nodes via DGX SuperPOD.
| GPU | 72x Blackwell Ultra GPU + 36x Grace CPU |
| CPU Core | 2,592 Arm Neoverse V2 cores |
| GPU VRAM | Bandwidth | 20 TB | Up to 576 TB/s |
| Total Quick Memory | 37 TB |
| FP4 Performance | 1440 PFLOPS (sparse) | 1080 PFLOPS (dense) |
| FP8/FP6 Performance | 720 PFLOPS (non-sparse) |
| Network Computing | 72x ConnectX-8 VPI (800 Gb/s InfiniBand) |
| NVLink Switch | 9x L1 NVLink Switch |
| Software | Mission Control / AI Enterprise / DGX OS |
| Support | 3-Year Business Standard Hardware and Software Support |
Brochures are generated in real-time from page content. If your browser blocks the download, allow pop-ups and try again.
Ready to deploy DGX GB300?
Huaying Technology provides end-to-end professional services for DGX GB300 configuration consulting, liquid-cooled rack planning, and deployment implementation.
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.