NVIDIA announced the DGX Spark, a desktop AI development system powered by the new GB10 Grace Blackwell superchip, capable of running models with up to 200 billion parameters locally. The system, priced between $2,999 and $3,999, is scheduled for delivery in Summer 2025.
The compact workstation targets developers and researchers who need local AI computing power without the recurring costs and data privacy concerns of cloud services, according to NVIDIA. The system delivers 1 petaFLOP of AI performance in a form factor smaller than most gaming consoles.
GB10 Grace Blackwell architecture details
The GB10 Grace Blackwell superchip integrates a 20-core ARM CPU with Blackwell generation GPU technology using NVIDIA's NVLink-C2C interconnect. The CPU combines 10 high-performance Cortex-X925 cores with 10 power-efficient Cortex-A725 cores, the company said.
The system includes 128GB of LPDDR5X unified memory with 273 GB/s bandwidth, enabling it to load large language models that would typically require multiple traditional GPUs. With fifth-generation Tensor Cores supporting FP4 precision, the system achieves its 1 petaFLOP performance while consuming 170 watts of power.
"This architecture eliminates traditional CPU-GPU bottlenecks through NVIDIA's NVLink-C2C interconnect, delivering 5x the bandwidth of PCIe Gen 5," the company announced.
Industry adoption and applications
Healthcare organizations are among early adopters, using DGX Spark to develop diagnostic AI models while maintaining HIPAA compliance, according to NVIDIA. A prominent medical research facility deployed multiple units to analyze patient imaging data without exposing sensitive information to cloud services.
Financial services firms reported significant cost reductions through local AI development. One major investment bank reduced AI development costs by 60% by moving initial model work from cloud services to DGX Spark clusters, NVIDIA stated.
Educational institutions have adopted the platform for advanced AI courses, giving students hands-on access to hardware capable of running state-of-the-art models, the company said.
Specifications and capabilities
The DGX Spark measures 150mm x 150mm x 50.5mm and weighs 1.2 kilograms. Connectivity includes four USB4 Type-C ports supporting up to 40 Gb/s data transfer and an integrated ConnectX-7 Smart NIC for 200GbE networking when clustering multiple units.
Two DGX Spark systems can work together on models up to 405 billion parameters, the company said. Storage options include 1TB or 4TB NVMe M.2 SSDs with hardware encryption. The system operates below 35 dB and powers through USB Type-C.
Software support includes DGX OS based on Ubuntu 22.04 with pre-installed PyTorch, TensorFlow and JAX frameworks optimized for ARM. The system also includes NVIDIA RAPIDS for data science workflows and access to the NGC catalog of pre-trained models.
Pricing and availability
The base configuration with 1TB storage costs $2,999, targeting budget-conscious developers and educational institutions. The 4TB Founders Edition at $3,999 targets professionals requiring additional storage for large datasets and model checkpoints, according to NVIDIA.
Operating costs at 170W power consumption translate to approximately $150-$200 annual electricity costs under typical development workloads. For organizations with consistent AI development needs, DGX Spark typically achieves return on investment within 4-6 months compared to cloud computing costs, the company stated.
Partner pricing varies by region, with ASUS offering the Ascent GX10 at $2,999 for similar specifications. European pricing runs higher at €3,689 due to import duties and VAT. Educational discounts through NVIDIA's academic program can reduce costs by 10-20%.
Models developed on DGX Spark can be containerized using the NVIDIA Container Toolkit and deployed directly to production environments without modification, preserving architectural consistency between development and deployment platforms.
