NVIDIA DGX Spark 64GB Overview

Image: NVIDIA · Source

The NVIDIA DGX Spark 64GB is a compact local AI supercomputer designed for running and scaling AI models on device, offering up to 100-billion-parameter model support and easy clustering for larger workloads.

The NVIDIA DGX Spark 64GB is a local AI platform poised to empower developers, researchers, and AI enthusiasts with a powerful, compact system optimized for running advanced AI models on device. As detailed by NVIDIA, this configuration is equipped with 64GB of unified memory and incorporates the GB10 Grace Blackwell Superchip, DGX OS, and the full NVIDIA AI software stack, including NVIDIA Agent Toolkit, CUDA-X AI libraries, and support for frameworks like PyTorch and popular runtimes such as Ollama and vLLM.

Aimed at supporting local AI workloads without reliance on cloud infrastructure, the DGX Spark enables users to run up to 100-billion-parameter models on a single device. When AI workloads exceed the capacity of a single unit, two DGX Spark 64GB systems can seamlessly cluster together using the NVIDIA Sync Cluster Assistant, pooling memory to 128GB and enhancing performance by up to 1.7 times in tested scenarios such as NVIDIA’s Qwen 3.8 27B model. This clustering is facilitated by built-in NVIDIA ConnectX-7 networking, allowing direct connection via QSFP cables.

NVIDIA emphasizes ease of use in setting up and scaling clusters, with automatic detection and configuration handled by the Sync app. The upcoming NVIDIA Sync Model Launcher aims to simplify running local AI models further by providing a GUI-driven experience for launching models like Qwen3.8 27B across single or multiple DGX Sparks.

Manufactured by partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI, the DGX Spark 64GB is offered as a more accessible option compared to higher-memory models, with a price starting at $4,999. The platform is suitable for tasks such as agent development, inference, fine-tuning, data science, and edge computing. Moreover, major creative applications like Blender have plans to support the platform, extending its usability.

This overview is based solely on NVIDIA’s published information and other source material and does not include hands-on testing or experiential insights.