NVIDIA DGX Station GB300
NVIDIA's reference GB300 workstation delivers 748 GB of coherent memory (252 GB HBM3e + 496 GB LPDDR5X), a 72-core Grace CPU, and 20,000 TFLOPS FP4 in a liquid-cooled 1,600W tower. The benchmark for desktop-class AI compute, capable of training and running trillion-parameter models locally.
Grace Blackwell Architecture · 748GB Unified Memory · 20.0K TOPS
Grace Blackwell Q1 2025 Unified Memory Ubuntu with NVIDIA AI Developer Tools
Product Image
Coming Soon
AI Performance
20.0K
TOPS
Memory
748GB
HBM3e
Max Model
1000B
parameters
Networking
800
Gbps
Full Specifications
| Architecture & Compute | |
|---|---|
| Architecture | Grace Blackwell |
| SoC Model | GB300 |
| GPU Architecture | NVIDIA Blackwell |
| Tensor Cores | 5th Generation |
| CPU Architecture | ARMv9 |
| CPU Cores | 72 cores |
| CPU Details | NVIDIA Grace 72-Core Neoverse V2 |
| Memory | |
| Capacity | 748 GB |
| Type | HBM3e |
| Bandwidth | 7100 GB/s |
| Interface | 256-bit |
| Unified Memory | Coherent CPU-GPU |
| Performance | |
| Peak TFLOPS | 80.0 |
| Peak TOPS (AI) | 20.0K TOPS |
| FP4 Performance | 20.0K TOPS |
| Max Model Parameters | 1000 billion |
| Networking & Connectivity | |
| SmartNIC | ConnectX-8 |
| NIC Bandwidth | 800 Gbps |
| Ethernet | 400GbE |
| USB Ports | Front: 2x USB-C 3.1, 2x USB-A 3.1; Rear: 4x USB-A, USB Micro-B (BMC) |
| Physical Specifications | |
| Dimensions (W×D×H) | N/A |
| Power Supply | 1600W |
| TDP | 1600W |
Similar Systems
NVIDIA DGX Spark Double Stack Bundle
Memory 256GB
Performance 2.0K TOPS
Architecture Grace Blackwell
NVIDIA DGX Spark Quad Stack Bundle
Memory 512GB
Performance 4.0K TOPS
Architecture Grace Blackwell
Release Information
| Release | Q1 2025 |
| Vendor | NVIDIA |
| OS | Ubuntu with NVIDIA AI Developer Tools |