Microsoft & NVIDIA Preview DGX Station for Windows With GB300 Blackwell Ultra Power
3 min. read 
Published on
Following the announcement in May earlier this year, NVIDIA has finally previewed its DGX Station platform for Windows in collaboration with Microsoft at today’s Surface & Windows event. Powered by the GB300 Grace Blackwell Ultra Desktop Superchip, the new DGX Station for Windows is designed to run frontier AI models and autonomous agents locally, reducing reliance on cloud infrastructure.
NVIDIA DGX Station Packs Up to 748GB of Memory
The headline number is hard to ignore. DGX Station for Windows can deliver up to 748GB of coherent memory and 20 PFLOPS of FP4 AI compute, giving it enough capacity to run AI models with up to 1 trillion parameters locally.
The system combines an NVIDIA Blackwell Ultra GPU with a 72-core NVIDIA Grace CPU, connected through NVLink-C2C. The GPU includes 252GB of HBM3e memory with 7.1TB/s bandwidth, while the system also has 496GB of LPDDR5X CPU memory.
NVIDIA is also allowing the system to be configured with an additional RTX PRO Blackwell workstation GPU. That gives developers a combination of large-scale AI compute and ray-traced graphics for visualization, simulation and physical AI workloads.
Windows Becomes Part of the DGX AI Stack
The biggest change is arguably the operating system. DGX Station previously focused on Linux, but the Windows version is designed to fit directly into enterprise environments already built around Microsoft’s platform.
Developers can build AI agents connected to Windows applications, run local inference and fine-tune models while continuing to use existing workflows. Linux-based AI frameworks can still be accessed through Windows Subsystem for Linux.
The machine also features NVIDIA’s AI software stack and ConnectX-8 SuperNIC, with networking speeds of up to 800Gb/s. Multiple DGX Stations can also be connected to scale model capacity and performance. NVIDIA says the system is aimed at AI-native companies, research teams and enterprises building long-running AI agents, data science workloads and frontier AI models.
DGX Station Technical Specifications
| NVIDIA GPU | 1x NVIDIA Blackwell Ultra |
| NVIDIA CPU | 1x Grace 72-Core Neoverse V2 |
| GPU Memory | 252 GB HBM3e | 7.1 TB/s |
| CPU Memory | 496 GB LPDDR5X | 396 GB/s |
| NVLink-C2C | 900 GB/s |
| Networking | Peak Bandwidth | NVIDIA ConnectX-8 SuperNIC | Up to 800 Gb/s | Ethernet |
| Supported OS | Microsoft Windows |
| FP4 Tensor Core | 20 | 153 PFLOPS |
| FP8 / FP6 Tensor Core | 10 PFLOPS |
| Int8 Tensor Core | 330 TOPS |
| FP16/BF16 Tensor Core | 5 PFLOPS |
| TF32 Tensor Core | 2.5 PFLOPS |
| FP32 | 80 TFLOPS |
| FP64 / FP64 Tensor Core | 1.3 TFLOPS |
| Total System Power | 1,600 W | 20A circuit required |
| Decoders | 7 NVDEC 7 nvJPEG |
| Supported RTX PRO GPU4 | NVIDIA RTX PRO 6000 Workstation Edition RTX PRO 6000 Blackwell Max-Q Workstation Edition RTX PRO 4000 Blackwell SFF Edition RTX PRO 2000 Blackwell |
| Storage | 4x M.2 Gen 5 Slots |
| Ethernet Ports | 2x QSFP112 (400 Gbs per port) 1x RJ45 10GbE 1x RJ45 1GbE (BMC for system management) |
| PCIe Slots | 1x PCIe Gen 5 x16 2x PCIe Gen 5 x16 (x8 electrical) |
| Additional Ports and Connectors4 | Front IO: 2x USB Type C (USB3.1), 2x USB Type A (USB3.1), Audio Rear IO: 4x USB Type A, USB Micro-B (BMC) mDP5 (BMC), Audio |
Coming to Windows Desktops in Q4
DGX Station for Windows is scheduled to arrive in Q4, although NVIDIA has not announced pricing yet. The system draws up to 1,600W and requires a 20A circuit, making it considerably closer to a compact data-center machine than a conventional desktop workstation.
Read our disclosure page to find out how can you help Windows Report sustain the editorial team. Read more
Improve this guide
User forum
0 messages