NVIDIA will offer the DGX Spark with 64GB of unified memory starting October 23rd through partners including Acer, ASUS, Dell, Gigabyte, HP and MSI. This new configuration allows developers to run up to 100-billion-parameter models fully on device, bypassing cloud reliance for local AI work. NVIDIA states that DGX Spark combines Grace Blackwell compute, unified memory and a CUDA-accelerated AI software stack into a complete local AI platform. Two DGX Spark units can also cluster together using NVIDIA Sync Cluster Assistant without additional setup, scaling projects beyond a single machine’s capacity.

DGX Spark 64GB Configuration and Partner Availability

The DGX Spark configuration now includes a 64GB unified memory option, becoming available through several major manufacturers on October 23rd. NVIDIA is streamlining the setup process for developers with DGX OS and the NVIDIA AI software stack pre-installed on the new configuration. The launcher will also automatically configure OpenCode, enabling developers to begin coding directly within their web browser.

This pre-configuration is intended to reduce friction and accelerate the development cycle for AI applications. The company highlights that the 64GB DGX Spark supports practical work from the outset, catering to both individual enthusiasts and larger research teams.

According to NVIDIA, the assistant configures the network and routes workloads automatically, eliminating the need for manual setup. This feature is particularly relevant for projects demanding more memory and compute power than a single machine can provide. The company states, “With up to 100-billion-parameter models running entirely on device, developers and enthusiasts can start with a single system for models that fit within its memory, or connect multiple DGX Spark systems.” Further resources, including agentic AI playbooks, are available on build.nvidia.com/spark.

NVIDIA Sync Cluster Assistant Enables Scalable Local AI

NVIDIA Sync Cluster Assistant streamlines expansion of local AI projects by automatically configuring networking and workload distribution between two DGX Spark units. The assistant eliminates the need for manual setup, allowing developers to pool memory resources and scale projects beyond the capacity of a single machine. This automated configuration is achieved via the 200 GbE fabric connecting the systems, enabling a combined 128GB of memory when two 64GB DGX Spark units are linked.

The ease of scaling is intended to support increasingly demanding AI workloads, such as running larger models or processing extended context windows. NVIDIA states that the same workflow developed for a single DGX Spark unit will function smoothly across two connected systems without requiring software reconfiguration. This feature addresses a key challenge in local AI development, where resource limitations can hinder experimentation and deployment of complex models.

The company highlights that this scaling capability allows developers to address tasks that exceed the memory and compute power of a single device, NVIDIA says. Beyond automated configuration, the DGX Spark platform also simplifies the software environment for developers.

64GB DGX Spark Supports 100-Billion-Parameter Models

Blender, a major creator application provider, will soon offer a prebuilt, downloadable installer to support the platform, demonstrating early industry adoption. In testing, two clustered 64GB systems delivered up to 1.7 times the performance compared with a single system when running NVIDIA’s Qwen 3.8 27B model, indicating a near-linear scaling benefit, according to the company.

The DGX Spark ships with support for several popular runtimes, including Ollama, vLLM and PyTorch with CUDA, alongside the NVIDIA Agent Toolkit and CUDA-X AI libraries. The 64GB configuration retains the GB10 Grace Blackwell Superchip, ensuring consistent performance with the 128GB model while offering a more accessible price point.

Source: https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/

More like this