MSI DGX Spark Blackwell 20 Core CPU +128GB GPU/4TB NVME SDD/3Y NVIDIA GB10

אודות המוצר

Download the presentation in PDF General Overview DGX Spark™ — a next-generation desktop system introduced by NVIDIA as a «mini-supercomputer for AI» on your desk. It is designed for those working with machine learning tasks, large models, inference, fine-tuning, and AI development — providing the power once reserved for data centers, now in a compact and convenient form factor. Thanks to the Grace Blackwell architecture (GB10 chip) and unified memory of 128 GB, DGX Spark™ can «run» large neural network models directly on a workstation or in a local office/lab environment. Key Technical Specifications and Features Architecture and Computational Power The system is based on a single-chip system (SoC) GB10 built on the Grace Blackwell architecture, combining a 20-core ARM processor (10 × Cortex-X925 + 10 × Cortex-A725) and a Blackwell-GPU accelerator. Tensor computation performance reaches up to 1 petaflop (1 PFLOP ) in FP4 mode (with sparsity optimization). Unified memory: 128 GB LPDDR5x with bandwidth up to 273 GB/s (256-bit interface) — accessible to both CPU and GPU without partitioning. Form factor: only 150 mm × 150 mm × 50.5 mm — a very compact system, weighing around 1.2 kg. Storage, Interfaces, and Connectivity Built-in NVMe M.2 SSD: 1 TB or 4 TB, with self-encryption support. Network interfaces: 10 Gb/s Ethernet (RJ-45), smart ConnectX-7 NIC, cluster support via 200 Gb/s RDMA (when connecting two units) for scaling to models with hundreds of billions of parameters. Wireless support: Wi-Fi 7, Bluetooth 5.x. Video and USB ports: HDMI 2.1a (for monitoring or output), 4× USB-C/USB4. Software and Ecosystem The system runs on DGX OS — NVIDIA’s AI-optimized operating system, with a preinstalled development stack (TensorFlow, PyTorch, and others). Development support: running and testing large language models, generation, training, inference, fine-tuning, remote access via NVIDIA Sync, and VS Code integration. Benefits for Developers, Researchers, and Businesses Work with very large models. With 128 GB of unified memory and shared CPU/GPU architecture, the system can load and serve models with hundreds of billions of parameters (in FP4 mode) — up to ≈ 200B per unit, and up to ≈ 405B when combining two DGX Sparks. Compactness and accessibility. Instead of a costly rack server with high power consumption, you get a desktop device that can fit in an office or lab — enabling local AI model development without reliance on cloud resources. Scalability when needed. If one device is not enough, two DGX Spark units can be clustered to work jointly with large models and distributed inference workloads. Flexible use cases. Suitable for fine-tuning custom models, inference, and research tasks — from text and image generation to scientific workloads — all locally, with lower latency and higher control. Integration with familiar tools. Thanks to popular framework and remote access support, developers can keep using their familiar environments while leveraging DGX Spark’s power in the background, improving efficiency and productivity. Reduced infrastructure costs. Using a compact system eliminates expenses for large server setups, rooms, and complex cooling — especially important for startups, labs, and small teams. Use Cases Development and testing of language models ( LLM ). Load models with hundreds of billions of parameters, experiment with architecture and hyperparameter tuning, and run inference services right in your office. Image, video, and media generation. The power of tensor cores and unified memory enables smooth generative modeling, pre/post-processing, and prototyping. Inference and AI deployment at the edge or in lab environments. DGX Spark can act as a «local AI server» for demos, pilots, and on-site implementations that require both power and autonomy. Scalable computing and cluster solutions. Combine multiple DGX Spark units for distributed computing, large data processing, and parallel multi-model workloads. Scientific research and computational tasks. The memory capacity and processing power make it appealing for researchers needing a local, controlled compute resource for experiments and algorithm development. Important Considerations Despite its desktop form factor, DGX Spark — is a specialized AI system, not a replacement for a regular personal or gaming PC. For rendering or gaming-only workloads, other solutions exist. Performance — though impressive, full-scale «training from scratch» for very large models (billions+ parameters) may still require higher-tier or distributed hardware. DGX Spark is optimal for development, fine-tuning, and inference of large models, but not always for «maximum training» like in data center clusters. Infrastructure and ecosystem: ensure adequate cooling, power, and network access, especially if clustering or intensive operation is planned. Price: while far more compact than data center alternatives, it remains a premium device targeted at AI development. Why DGX Spark — A Step Forward Before such systems existed, AI developers often faced a dilemma: run models in the cloud (leading to costs, latency, and external dependency) or use local PCs with memory and performance limitations. DGX Spark changes that: the computing power once confined to server clusters can now sit right on your desk. The unified memory architecture between CPU and GPU eliminates data transfer overhead between subsystems, allowing work with large models without complex setup. Clustering multiple devices takes it further: you can start with one unit and expand later as needed. Moreover, NVIDIA’s ecosystem — optimized drivers, libraries, development tools, and framework support — reduces deployment time. Researchers, developers, and businesses gain a tool that accelerates the path from idea to prototype to production. Conclusion If you are a developer, researcher, or AI engineer seeking a device that provides maximum local capability without immediately building a server infrastructure, DGX Spark™ is one of the best options available: Large unified memory (128 GB) shared between CPU and GPU. Impressive computational performance (up to 1 PFLOP) powered by Grace Blackwell architecture. Compact form factor suitable for office or lab environments. Flexibility: from development and fine-tuning to cluster deployments. Full integration with the NVIDIA ecosystem and support for popular frameworks.

מפרט טכני

• Manufacturer Product Link Presentation
• DGX OS with NVIDIA’s AI software stack
• ARM
• 128 GB
• 4 TB
• Warranty: 36 Month
• Warranty type: Standard