NVIDIA DGX™ GB300
NVIDIA DGX GB300 is a rack-scale, liquid-cooled AI system built for enterprise AI factories and the era of AI reasoning. Built with NVIDIA Grace™ Blackwell Ultra Superchips, it is designed to maximize performance for training, post-training, and test-time inference workloads.
- 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace™ CPUs
- 20 TB total GPU memory with up to 576 TB/s memory bandwidth
- Up to 1,440 petaFLOPS of FP4 Tensor Core performance
- 2,592 Arm Neoverse V2 CPU cores
DGX GB300 Technical Specifications
| Specification | Details |
|---|---|
| GPU |
|
| CPU Cores |
|
| Total GPU Memory | Bandwidth |
|
| Total Fast Memory |
|
| Performance |
|
| Networking |
|
| NVIDIA NVLink Switch System |
|
| Management Network |
|
| Software |
|
| Enterprise Support |
|
* *Specification shown in sparse | dense
