NVIDIA DGX Station GB300

The NVIDIA DGX Station GB300 is a desktop AI supercomputer built on the Grace Blackwell Ultra architecture, combining NVIDIA's GB300 Blackwell Ultra GPU with Grace Arm-based CPU cores in a tightly integrated NVLink-C2C interconnect fabric. This coherent memory architecture eliminates traditional PCIe bandwidth bottlenecks, enabling the CPU and GPU to share a unified 784 GB memory pool at full NVLink bandwidth — a capability previously available only in large-scale data center nodes. The result is a workstation that can hold entire frontier-scale large language models and multimodal AI models entirely in memory, enabling iterative fine-tuning, inference, and agentic AI workflows without the latency of model sharding across multiple nodes.

The DGX Station GB300 targets AI researchers, computational scientists, and enterprise development teams who require data center-class AI compute at their desk without dependence on shared cloud or cluster resources. Its unified memory pool is particularly transformative for workflows involving 70B to 400B+ parameter models, genomics pipelines, physics-based simulations, and real-time multimodal inference that demand both massive memory capacity and high memory bandwidth simultaneously. The integrated Grace CPU cores provide substantial general-purpose throughput for data preprocessing, orchestration, and hybrid CPU-GPU workloads without requiring a separate host processor platform.

As part of the NVIDIA DGX ecosystem, the GB300 Station ships with DGX OS and the full NVIDIA AI Enterprise software stack, including optimized containers for NeMo, TensorRT-LLM, cuDNN, and RAPIDS. This software integration ensures teams can transition workloads seamlessly between the DGX Station and DGX SuperPOD clusters. The system is engineered for enterprise deployment with managed cooling, NVMe storage, and high-bandwidth networking in a form factor designed for office or lab environments, making it the highest-density single-node AI compute platform available outside a data center rack.

Ideal for

• On-premises fine-tuning and alignment of 70B to 400B+ parameter large language models without model sharding or multi-node orchestration overhead
• Real-time multimodal AI inference combining vision, language, and structured data models held entirely within the 784 GB unified memory pool
• Computational genomics and drug discovery pipelines requiring simultaneous large-memory graph analytics and GPU-accelerated molecular simulation
• Agentic AI development and testing where researchers iterate rapidly on autonomous reasoning systems requiring both high compute throughput and large context windows
• Digital twin simulation and physics-based AI modeling for engineering, manufacturing, and autonomous systems development at the researcher's workstation
• Enterprise AI application prototyping that mirrors production DGX SuperPOD workloads, enabling validated model development before cluster-scale deployment

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.