NVIDIA Expands Personal AI Supercomputer Lineup, Unveils DGX Spark 64GB

NVIDIA said on the 6th that it is expanding its lineup of personal AI supercomputers and unveiled the 64GB configuration of DGX Spark, a compact personal AI supercomputer. The product is aimed at developers, researchers, and AI enthusiasts.

DGX Spark 64GB is designed to support the development and execution of large AI models and agents in a cloud-independent local environment. At the same time, it offers the same GB10 Grace Blackwell Superchip as the existing 128GB model, the same DGX OS, and the same NVIDIA AI software stack, while aiming to lower the barrier to adoption.

DGX Spark 64GB will launch in October and will be distributed through manufacturing partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI.

DGX Spark can run models of up to 100 billion parameters and agent applications built on them directly on a single device. A key feature of DGX Spark 64GB is that it is not limited to a single system and can scale through multi-system connections.

DGX Spark combines Grace Blackwell computing, unified memory, NVIDIA ConnectX-7 networking, and the CUDA-accelerated AI software stack into a single system. The device supports agent development, inference, fine-tuning, data science, and edge development workloads in local AI environments.

Every DGX Spark comes standard with a ConnectX-7 NIC. Two DGX Sparks can be directly connected with a QSFP cable to form a cluster, in which case total memory expands to 128GB. As a result, support extends to models of up to 200 billion parameters.

On the performance side, memory bandwidth doubles. In addition, based on NVIDIA's own tests, the Qwen 3.8 27B model delivers up to 1.7 times the performance of a single-system setup when configured as two systems.

To make this expanded configuration easy to deploy, NVIDIA Sync Cluster Assistant simplifies multi-node setup, automatically detects connected systems, verifies device configuration, and handles ConnectX-7 network configuration. Each system uses the same NVIDIA software stack, so no separate software reconfiguration is needed when expanding from a single system to a two-unit cluster.

Separately, DGX Spark provides a software environment for agent development. Supported software includes NVIDIA Agent Toolkit, CUDA-X AI libraries, and the Nemotron open models. Supported runtimes include Ollama, vLLM, and CUDA-enabled PyTorch.

DGX Spark can run models within minutes after the system powers on. Developers can experiment with models locally without sending data to an external cloud.

Its use cases are also expanding into the creator space. Blender is included among the major creator applications supported by DGX Spark, and a preconfigured installer for Blender will also be provided.

NVIDIA will offer Sync Model Launcher beginning later this month. With it, developers will be able to download and run the Qwen 3.8 27B model on a single DGX Spark or a cluster with just a few clicks.

Sync is configured so that models run on connected devices. As a result, the model can be accessed from a laptop, OpenCode can also be configured to use the model, and developers can start coding directly in a browser.

Through DGX Spark 64GB, NVIDIA is presenting a strategy to expand personal AI computing environments, saying that the expansion is moving beyond the scope of a simple local inference device toward an environment that can also support agent development, AI application building, and model scaling.

Source: TECHWORLD · Park Kyu-chan
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407799

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407799