NVIDIA DGX Station for Windows Deskside Supercomputer Local AI Agent Execution Blackwell Ultra Hardware Specs and Secure Enterprise Workload Scenarios
NVIDIA has unveiled the DGX Station for Windows a deskside supercomputer designed for the execution of local AI agents and large neural networks. The system can operate on frontier models with a trillion parameters directly on local workstations. This design combines datacenter class infrastructure with desktop components, empowering enterprises to run demanding development workloads locally rather than exclusively via remote Linux systems.
Previously, machine learning tasks such as model training and inference on a massive scale were performed using datacenters equipped with Linux. This innovation in the DGX Station for Windows brings the GB300 Grace Blackwell class hardware to the standard Windows operating system.
According to Chris Marriott, VP of enterprise platforms at NVIDIA, organizations need to run large AI agents locally and establish hardware connections to the productivity tools and applications used on a daily basis as their needs expand.
The DGX Station uses the GB300 Grace Blackwell Ultra Desktop Superchip. This single component connects a Blackwell Ultra GPU to a 72 core Grace CPU, taking full advantage of the NVLink C2C interconnect. The system includes up to 748GB of coherent memory and offers a computing power of 20 petaflops at FP4. With an added RTX PRO 6000 Blackwell GPU, the workstation can be used for the complex physics simulation and ray traced visualization demanded by the latest engineered products.
The ConnectX 8 SuperNIC handles high speed clustering and networking capabilities, offering a transfer speed of 800Gb/s which speeds up the intake of large datasets and permits multiple DGX Stations to be physically linked for work.
Pavan Davuluri, executive VP of Windows and Devices at Microsoft, stated that their partnership scales the full capabilities of Windows from slim devices to workstations and is enabling new levels of localized computing.
The DGX Station can also act as a separate workstation on which users can build, test and run AI agents that continuously run in a secured runtime environment. NVIDIA OpenShell a secure by design open source runtime environment is used to build and test the agents. OpenShell leverages Windows security features and containment primitives for each individual agent running in a sandbox. This feature ensures that agents will not change system policies, access invalid credentials, or leak confidential information in the process of running their tasks.
The workstation can execute any range of engineering and design work tasks in an enterprise. Developers will be able to run their familiar Linux AI toolchains on their Windows machine by utilizing the Windows Subsystem for Linux, while still gaining the benefits of using their standard Windows productivity suite. The 748GB of unified memory on the system ensures data science workloads can be scaled effectively and do not encounter performance issues when data is processed from large scale datasets. When coupled with the GB300, the addition of an RTX PRO 6000 GPU can be utilized for physical AI tasks, thereby integrating language modeling with real time spatial simulation.
The DGX Station for Windows can be either an individual workstation for one developer or a shareable local resource within an entire team. This device can be acquired from ASUS, Dell Technologies, GIGABYTE, HP, MSI, and Supermicro. The supercomputers are expected to be available in Q4.
