NVIDIA DGX Spark

The NVIDIA DGX Spark is built around the GB10 Grace Blackwell Superchip, integrating a Blackwell GPU with 128GB of unified LPDDR5X memory shared between the CPU and GPU subsystems via a high-bandwidth NVLink-C2C interconnect. This architecture eliminates PCIe bottlenecks between processor and accelerator, enabling memory bandwidth that far exceeds traditional discrete GPU workstations and allowing models up to 200 billion parameters to run entirely in local memory. The system measures roughly the size of a standard hardcover book, yet delivers computational density previously associated with full-rack AI infrastructure, making it the first desktop-class system to reach the 1 petaFLOP FP4 inference milestone.

DGX Spark targets AI researchers, data scientists, MLOps engineers, and application developers who require on-premises inference and fine-tuning capabilities without cloud dependency or data-egress constraints. The platform ships with NVIDIA AI software including NIM microservices, NeMo frameworks, and CUDA libraries, providing a fully validated software stack that mirrors the DGX Cloud environment. This continuity allows teams to develop and test locally on DGX Spark then scale workloads to DGX Cloud or DGX SuperPOD infrastructure with zero code changes, accelerating the iteration cycle from prototype to production.

From a connectivity standpoint, DGX Spark provides Thunderbolt 5 and USB4 ports, enabling it to be paired with NVIDIA DGX Station or external storage enclosures for dataset-intensive workflows. Its fanless-adjacent thermal design and quiet operation make it suitable for deployment in office environments alongside standard workstations. Running on standard AC power without special electrical infrastructure, DGX Spark democratizes petaFLOP-class AI compute for organizations that previously could not justify a dedicated AI server room, while still supporting multi-node scaling when multiple units are networked together.

Ideal for

• Local inference of large language models up to 200B parameters for enterprise chatbots and coding assistants without sending sensitive data to external cloud APIs
• Rapid fine-tuning and LoRA adaptation of foundation models on proprietary datasets under strict data-governance or regulatory requirements
• Edge AI development and validation for robotics, autonomous systems, and industrial computer vision pipelines before deployment on NVIDIA Jetson or embedded platforms
• Research prototyping of multimodal AI applications combining vision, language, and structured data in an isolated, reproducible on-premises environment
• MLOps teams running continuous evaluation benchmarks and regression testing of model checkpoints locally to reduce cloud compute expenditure during the iterative training loop
• Small AI-focused teams at startups or enterprise innovation labs requiring a shared, always-available AI compute node without the overhead of rack-mount server provisioning

Technical specifications

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