NVIDIA DGX Spark: Everything You Need to Know
From initial setup to hands-on testing to our first impressions of this deskside AI machine.News
It's been just over a week since the NVIDIA DGX Spark officially launched, transitioning from the "Project Digits" concept we saw at CES to a real product available for purchase at your local Micro Center. In that time, we've had a chance to do some initial hands-on testing and dig into what it’s actually like to live and work with this compact AI supercomputer.
But at the same time, our initial testing has barely scratched the surface of what this machine can do. We’ve got plenty of plans for more DGX Spark content, so be sure to keep an eye on MC News and our YouTube page.
What's Inside the DGX Spark?
The NVIDIA DGX Spark is built around the Grace Blackwell architecture, a system-on-a-chip (SoC) design that is key to its performance. It features a highly integrated configuration with a 20-core Arm processor (combining 10 Cortex-X925 performance cores with 10 Cortex-A725 efficiency cores) and a Blackwell architecture GPU (with 6,144 CUDA cores) on the same package. This tight integration is crucial as it minimizes the latency that normally occurs when data has to travel between a separate CPU and GPU. For AI workloads that are constantly moving massive datasets, this design eliminates a major performance bottleneck. On the raw performance side, the system is rated for up to 1 PetaFLOP of AI performance at FP4 precision, a number that was reserved until recently for massive data center servers.
Photo: Jacob Bobo The most significant spec is arguably the 128GB of LPDDR5x unified system memory, which is accessible across a 256-bit interface, providing 273 GB/s of bandwidth that both the CPU and GPU share dynamically. This architecture is key for AI development on the desktop, as it allows the DGX Spark to handle incredibly large AI models of up to 200 billion parameters on a single system, without the complex memory management required on less capable systems.
Beyond the core processing, the DGX Spark's connectivity and physical design help cement it as a serious development tool. It's packed with useful I/O, including Wi-Fi 7, 10GbE Ethernet, four USB Type-C ports, and an HDMI 2.1a port. Most importantly, it includes a ConnectX-7 Smart NIC with two QSFP network connectors, providing up to 200 Gbps of bandwidth. This allows you to link two DGX Spark systems together using a single cable, creating a virtual cluster. This effectively creates a small, two-node supercomputer on your desk, enabling you to tackle even larger models (up to 405B parameters)
Packing this level of hardware into a tiny 1.2 kg chassis that draws only 240W of power is a massive engineering feat, especially when compared to the 3,200W power consumption of the original DGX-1 from 2016.
Who is the DGX Spark for?
The NVIDIA DGX Spark is for those that require a local, high-performance environment for AI development and experimentation. This includes AI/ML developers, data scientists, and academic researchers who need to prototype, debug, and validate complex models without the high costs of cloud-based GPUs. For example, someone who works with Integrated Development Environments like VS Code or Cursor and wants to use the NVIDIA Sync tool to connect local editors to the DGX Spark.
This platform is designed to bridge the gap between initial experimentation on a commercial laptop or desktop and expensive large-scale training in a data center by offering a compact, affordable deskside device that's compatible with common cloud and server architectures (like Blackwell).
The suggested use cases include computationally intensive AI workloads that would benefit from the unified memory architecture and Blackwell's Tensor Cores. It's especially suited to something like fine-tuning a model like Llama 3 8B through LoRA (low-rank adaptation). For inference, the system leverages its FP4 precision and speculative decoding to accelerate LLM performance, enabling the local deployment of complex, multimodal applications like RAG (Retrieval-Augmented Generation) agents. It can do generative media as well, running Flux AI in ComfyUI to create high-resolution 1K images, thanks to the large, unified memory pool.
How to use the DGX Spark: Desktop vs. Network modes
During the initial setup utility, you'll choose between two primary modes. Don't worry, you can switch between these modes freely after the initial setup is complete.
Desktop Mode: This is the traditional setup. You connect a monitor, keyboard, and mouse and interact directly with the Ubuntu-based desktop environment, called the DGX Dashboard.
Network Appliance Mode: In this mode, the Spark acts as a headless server on your network. You access it remotely from another computer, which is perfect for developers who want to connect their local IDE (like VS Code or Cursor) to the Spark using the NVIDIA Sync tool.
How Does the DGX Spark Perform?
Before talking about performance, it’s important to understand the two main jobs in the world of AI: Inference and Training.
Inference is the part you're likely familiar with. It's the act of using a trained AI to get results -- asking ChatGPT a question, having Copilot write code, or generating an image with Gemini. Think of it as a consumer-facing finished product.
But training and fine-tuning are how the AI model gets to that point in the first place. This is, essentially, the research and development phase. Training a huge model from scratch is a massive job that happens in data centers. Fine-tuning, however, is like adding custom functions to an AI for a specific task.
The DGX Spark is built for developers who do the fine-tuning of a model, then run inference on them locally, making it a complete, self-contained AI environment for creating, testing, and perfecting AI tools.
The Blackwell GPU is designed with exactly that kind of quantizing and fine tuning in mind. Our team quantized a Deepseek R1 Distillation of Llama 3 8B, after downloading necessary files first, in approximately 2:30. After that, it took the DGX Spark approximately 36 minutes to fine tune an 8B Llama Factory model provided to us by NVIDIA.
But note that, compared to a computer built for inference -- for example, our custom desktop with an AMD Threadripper and NVIDIA RTX Pro 6000 that costs quite a bit more than the Spark -- it's not going to return results as quickly. The Threadripper system responded to our test query with a speed of 28.29 tokens per second (TPS), while the DGX Spark scored 4.21 tokens per second.
The takeaway is that the DGX Spark is designed for developers, researchers, and enthusiasts who want to train or fine-tune their own models locally. This unique architecture is what allows for longer sequence lengths and higher-resolution images during training, effectively bringing a datacenter-level capability to your desk.
We’ll have plenty of additional benchmarks, guides, and tips for the DGX Spark coming to both Micro Center News and our YouTube channel, so keep checking back. If you want to get your hands on your own DGX Spark, it is available today, October 15th, chain wide, so swing by your local Micro Center and get ready to do AI work the supercomputer way.
MC News associate editor Sean Mekinda and editor-in-chief Dan Ackerman contributed to this report. Originally published October 13, 2025. Updated October 21, 2025 with new details and hands-on testing.
More from MC News
- How to Build a PC with a Hardline Water-Cooling Loop
- 3D Print a Mac Mini Monitor Mount
- Windows 10 End of Support Countdown: Less than 50 Days Left to Upgrade
- Everything You Need to Know About WiFi 7
- Keyboard 101: Intro to Computer Keyboards
- The NVIDIA 50 Series Laptop Buying Guide
- Can Your PC Run OpenAI's New GPT-OSS Large Language Models?
