NVIDIA has unveiled a Windows-compatible version of its DGX Station, expanding access to its Blackwell Ultra AI platform in a workstation form factor. The system, centered on the GB300 "Blackwell Ultra" GPU, delivers 20 PFLOPs of AI compute and supports up to 748 GB of unified memory.

Previously, the DGX Station operated exclusively on Linux, requiring enterprise developers to maintain separate Linux environments for AI workloads and Windows systems for productivity tools. This dual-environment model created inefficiencies, particularly for Fortune 500 companies standardizing on Windows. The new Windows variant aims to streamline workflows by consolidating AI development and application use within a single Windows-based platform.

The GB300 GPU features 160 streaming multiprocessors with 128 CUDA cores each, totaling 20,480 CUDA cores and 640 Tensor cores. It includes 40 MB of Tensor memory and supports FP8, FP6, and NVFP4 precision standards, achieving near-FP8 accuracy with minimal deviation. Memory capacity reaches 252 GB of HBM3e with 7.1 TB/s bandwidth, enabling a 50% increase in dense low-precision compute performance compared to prior Blackwell models. The system also integrates a 72-core Grace CPU based on the Neoverse V2 architecture and 496 GB of LPDDR5X system memory, connected via a 900 GB/s NVLink-C2C interconnect.

Combined, the CPU and GPU form a 748 GB unified memory pool, enhancing data throughput for large-scale AI models. Networking capabilities include an NVIDIA ConnectX-8 SuperNIC with 800 Gb/s speeds and support for multiple high-bandwidth Ethernet and PCIe Gen5 expansion slots. The workstation is rated at 1,600W TDP and will launch in Q4 of this year, with an estimated retail price near $100,000.

The announcement follows AMD's introduction of its Threadripper Halo Station, a competing high-end workstation platform combining Zen 5 CPUs and Radeon PRO GPUs, also targeting AI and professional workloads. While AMD emphasizes expandable memory configurations and ROCm ecosystem compatibility, NVIDIA highlights superior AI compute density and software integration through CUDA and TensorRT. Both platforms reflect growing demand for localized AI development tools as model complexity increases.