NVIDIA DGX Spark is designed as a personal AI workstation for data scientists, researchers, and developers. It is primarily used for AI model training, deep learning research, and high-performance computing tasks, providing enterprise-level AI capabilities on a desktop-scale system.

DGX Spark delivers exceptional performance for AI training workloads. Equipped with multiple GPUs, high-speed memory, and advanced networking, it can handle large datasets and complex models efficiently, drastically reducing training time compared to traditional workstations.

Yes. NVIDIA DGX Spark is optimised for training and fine-tuning large language models (LLMs). Its GPU acceleration, high memory bandwidth, and support for mixed-precision computing make it ideal for modern NLP workloads.

Preliminary, subject to change:

• Blackwell GPU - Up to 1,000 AI TOPS FP4
• Tensor Cores: 5th Generation
• RT Cores: 4th Generation
• CUDA Cores: Blackwell Generation
• Grace CPU: 10 performance cores (P-Cores) Cortex-X925 + 10 efficiency cores (E-Cores) Cortex-A725
• NVLink-C2C (chip to chip interconnect) 5x the bandwidth of PCIe 5
• Memory: 128GB of coherent, unified LPDDR5X Low-Power Memory
• Memory bandwidth: 273 GB/s
• Storage: Up to 4TB SSD storage
• Network connectivity: 10-GbE (RJ-45) and Connect-X (x2 QSFP)
• Display: HDMI 2.1
• NIC: ConnectX-7 SmartNIC
• USB: 4x USB Type C
• WiFi: WiFi-7
• Bluetooth: BT 5.3
• Audio output: HDMI multichannel audio output
• NVENC | NVDEC: 1x | 1x
• OS: NVIDIA DGX OS
• System dimensions: 150mm L x 150mm W x 50.5mm H
• Weight: 1.2kg

The GB10 Superchip is a system-on-a-chip based on a Grace Blackwell architecture with 128GB of unified memory. It provides up to one petaflop of AI performance at FP4 precision and features a NVIDIA Blackwell GPU with the latest fifth-generation Tensor Core and RT Core technology, along with a high-performance Grace CPU built with 20 Arm cores.

FP4 (Floating Point 4-bit) precision is a new AI processing format that allows models to use lower-precision arithmetic without compromising accuracy. This increases computational efficiency, enabling faster model training and inference on DGX Spark.

ConnectX-7 networking provides ultra-low latency and high throughput, which is crucial for AI workloads involving large datasets or distributed computing. It ensures faster communication between GPUs and other system components, maximising training efficiency.

A personal AI workstation like DGX Spark offers immediate access to powerful AI resources, allowing data scientists to train, test, and deploy models independently. It reduces reliance on cloud resources, enhances data privacy, and provides a seamless workflow for AI development.

DGX Spark supports popular AI frameworks and libraries, such as PyTorch and TensorFlow. You can deploy models directly from the workstation or integrate with NVIDIA AI Enterprise software to manage production-scale AI workloads efficiently.

The GB10 architecture allows enterprises and researchers to prototype, fine-tune, and test models on local DGX Spark systems and deploy in a data center or cloud. This enables efficient fine-tuning and inference workloads while maintaining consistency across development and deployment environments.

GB10 has ten Cortex-X925 and ten Cortex-A725 cores.

DGX Spark ships with DGX OS and additional NVIDIA tools. DGX OS is based on Ubuntu and is consistent with the Linux distribution running on other DGX platforms.

DGX Spark comes pre-installed and/or supports a large catalog of tools, frameworks, models and software for development use.

DGX Spark users can register free of charge for the NVIDIA Developer Program and NVIDIA NGC to get access to SDKs, toolkits, trained models, workflows, and much more.

Users can join the NVIDIA Developer Program and get access to extensive software support, documentation and the developer community for rich community-based support.

Hardware support will be provided by the vendor.

Customers are welcome to choose between the options available for sale. We do not support upgrades to DGX Spark at this time. Any attempt to tamper with or change the components will void the warranty.

We support connecting of up to 2 DGX Spark with an approved QSFP cable at this time.