NVIDIA DGX Spark

Description: AI supercomputer with Grace Blackwell on your desk – up to 1 PetaFLOP AI performance

Website: https://www.nvidia.com/en-us/products/workstations/dgx-spark/

NVIDIA DGX Spark is a compact desktop AI supercomputer with the GB10 Grace Blackwell Superchip. With up to 1 PetaFLOP FP4 AI performance and 128 GB unified memory, developers can prototype, fine-tune and infer AI models with up to 200 billion parameters locally.

Technical Highlights

• GPU/CPU: NVIDIA Grace Blackwell architecture with 20 Arm compute units (10x Cortex-X925 + 10x Cortex-A725)
• AI Performance: Up to 1 PetaFLOP at FP4 precision
• Memory: 128 GB LPDDR5x coherent unified system memory (273 GB/s bandwidth)
• Storage: 4 TB NVMe M.2 SSD with self-encryption
• Network: ConnectX-7 Smart NIC, 10 GbE, WiFi 7
• Power: 240W power supply, 140W TDP (GB10)
• Form Factor: 150 x 150 x 50.5 mm, 1.2 kg

Use Cases

Prototyping & Development: Development, testing and validation of AI models with pre-installed NVIDIA AI software stack.

Fine-Tuning: Optimization of pre-trained models with up to 70 billion parameters thanks to large unified memory.

Inference: Execution of cutting-edge AI models with up to 200 billion parameters thanks to 128 GB memory and 1 PFLOP compute power.

Edge Applications: Development of robotics, smart cities and computer vision with NVIDIA Isaac, Metropolis and Holoscan.

Scalability: Two DGX Spark systems can be connected via ConnectX networking to work with models up to 405 billion parameters.

Software

Pre-installed NVIDIA AI software stack with tools, frameworks, libraries and pre-trained models. Supports current models from DeepSeek, Meta, NVIDIA, Google, Qwen and others. Runs on NVIDIA DGX OS.

Availability

Available through NVIDIA marketplace and authorized resellers. OEM partners also offer GB10-based personal AI supercomputers.