Nvidia DGX Spark is a small-form-factor desktop computer developed by Nvidia for local artificial intelligence development and deployment. It combines an Arm-based processor, a Blackwell graphics processor, coherent unified memory, high-speed networking, and Nvidia’s AI software environment. Introduced under the name Project DIGITS in January 2025 and subsequently renamed, it became available to order on October 15, 2025. Nvidia markets the system as a “personal AI supercomputer,” principally for prototyping, inference, and fine-tuning rather than as a replacement for large data-center training systems. (nvidianews.nvidia.com)
Development and release
Nvidia announced Project DIGITS on January 6, 2025. The proposed desktop system used the GB10 Grace Blackwell Superchip and was intended to let researchers, developers, and students experiment with large models locally before transferring their work to accelerated servers or cloud computing services. Nvidia stated that MediaTek collaborated on the GB10’s design. (nvidianews.nvidia.com)
On March 18, 2025, Nvidia announced the DGX Spark name alongside the larger DGX Station. Although both belonged to its desktop AI computing initiative, they were distinct products: Spark used GB10, whereas the announced Station used the Blackwell Ultra platform. Nvidia announced Spark’s shipment arrangements on October 13, with ordering beginning on October 15, 2025. (investor.nvidia.com)
The original configuration provided 128 GB of unified memory. On October 2, 2026, Nvidia announced a 64 GB configuration through participating computer manufacturers, with availability scheduled for October 23, 2026. As of October 15, 2026, that configuration had been announced but had not reached its scheduled release date. (blogs.nvidia.com)
Hardware architecture
The GB10 combines a 20-core central processing unit with a Blackwell graphics processing unit. The CPU implements the Arm architecture and contains ten Cortex-X925 cores and ten Cortex-A725 cores. The GPU includes fifth-generation Tensor Cores, specialized hardware for the matrix operations used extensively in deep learning. (docs.nvidia.com)
A defining feature is coherent unified memory, shared by the CPU and GPU. Nvidia’s NVLink-C2C interconnect maintains coherence between the processing components. Unlike a conventional arrangement with separate CPU memory and discrete GPU memory, this design provides a common memory pool for CPU–GPU workloads. (nvidianews.nvidia.com)
Key specifications of the original 128 GB system include:
| Component | Specification |
|---|---|
| Processor | Nvidia GB10 Grace Blackwell Superchip |
| CPU | 20 Arm cores |
| GPU | Blackwell architecture; 6,144 CUDA cores |
| Memory | 128 GB LPDDR5X unified memory |
| Memory interface and bandwidth | 256-bit; 273 GB/s |
| Storage | NVMe M.2 solid-state storage, with configurations up to 4 TB |
| Networking | 10 GbE, ConnectX-7 networking rated at 200 Gb/s, Wi-Fi 7 |
| External connections | Four USB-C ports and HDMI 2.1a |
| Reference enclosure | 150 × 150 × 50.5 mm; approximately 1.2 kg |
| Power | 240 W external supply; 140 W GB10 thermal design power |
The power-supply rating and processor thermal design power describe different quantities; neither is a statement that the complete system continuously consumes that amount of electricity. (docs.nvidia.com)
Software and workflows
DGX Spark runs Nvidia DGX OS, an Ubuntu-based Linux operating system with hardware-specific drivers, configuration, diagnostics, and maintenance tools. Its software environment connects the desktop hardware to Nvidia’s CUDA ecosystem and associated libraries and frameworks for machine learning. (docs.nvidia.com)
Typical workflows include running large language models, testing applications, and adapting pretrained models before deployment elsewhere. At launch, Nvidia described the 128 GB system as supporting inference with models of up to 200 billion parameters and fine-tuning of models of up to 70 billion parameters. These are workload-dependent capability claims, not guarantees for every model or training method. (nvidianews.nvidia.com)
Local deployment also supports applications in which model execution takes place on a nearby dedicated computer rather than through a remote inference service. Nvidia subsequently expanded its Spark software and documentation toward AI agents, including development environments and multi-agent workloads. (blogs.nvidia.com)
Networking and multi-system operation
ConnectX-7 networking allows Spark units to participate in distributed computing workloads. Nvidia’s later software supports connecting two to four systems through the NVIDIA Sync Cluster Assistant. Four 128 GB units provide 512 GB of aggregate memory capacity for supported distributed applications. This expands the available resources across multiple computers rather than replacing their network connection with a single machine’s internal memory interface. (developer.nvidia.com)
Clustering can accommodate larger models and additional concurrent work, but the benefit depends on the software and workload. Nvidia’s published performance discussions distinguish local memory bandwidth, computation, and inter-node communication rather than treating additional systems as an automatic proportional speed increase. (developer.nvidia.com)
Performance claims and limitations
The advertised one petaflop figure is a theoretical peak for four-bit floating-point Tensor Core operations using sparsity. It is not a general-purpose performance rating at FP32 or FP64 precision, nor a direct measure of language-model generation speed. Comparisons require matching numerical formats, sparsity assumptions, and workloads. (nvidia.com)
Memory capacity and memory bandwidth also serve different purposes. A large shared memory pool can make a model fit, while the rate at which its data can be accessed still constrains execution. Nvidia’s technical analyses distinguish compute-bound and bandwidth-bound workloads and show that performance also depends on kernel optimization, model configuration, and inference software. Consequently, the system’s model-size limits and peak arithmetic rating cannot by themselves predict application latency or throughput. (developer.nvidia.com)
References
- NVIDIA Puts Grace Blackwell on Every Desk and at Every AI Developer’s Fingertipsnvidianews.nvidia.com
- NVIDIA DGX Spark Arrives for World’s AI Developersnvidianews.nvidia.com
- NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AIblogs.nvidia.com
- Personal AI Supercomputer Powered by Blackwellnvidia.com
- Hardware Overview — DGX Spark User Guidedocs.nvidia.com
- DGX Spark User Guidedocs.nvidia.com
- DGX OS — DGX Spark User Guidedocs.nvidia.com
- DGX Spark Software Stack — NVIDIA DGX Spark Porting Guidedocs.nvidia.com
- Run Local AI Agents with Faster Models and Multi-Node Clustering on NVIDIA DGX Sparkdeveloper.nvidia.com
- Scaling Autonomous AI Agents and Workloads with NVIDIA DGX Sparkdeveloper.nvidia.com
- Expert Assistance for AI Infrastructure — NVIDIA DGX Spark Supportnvidia.com