Overview
- Authors: Christof Windeck, Jan-Keno Janssen
- Source: https://www.heise.de/-10867427
- Date: 14.11.2025
- Estimated reading time: 12 minutes
Summary
The Nvidia DGX Spark combines mini PC, Linux workstation and AI server in one device and is specifically aimed at AI developers who want to run local language models.
- 128 GByte LPDDR5X RAM with 273 GByte/s transfer rate enables local operation of large AI models
- GB10 combo processor with 20 ARM cores and CUDA-compatible GPU with 6144 shader units
- Price of Gigabyte version: around 4300 Euro, also available from Asus, Dell, HP and Lenovo
- 200 Gbit/s network adapter ConnectX-7 for connecting multiple devices
- Pre-installed DGX OS 7 (Ubuntu 24.04.03 LTS) with AI development tools
- Performance: 45 tokens/s with GPT-OSS:120b, 216 watts maximum power consumption
- Only one HDMI output, four USB-C ports without Thunderbolt or USB4
Opportunities & Risks
Opportunities
- Local AI development without cloud dependency
- CUDA compatibility for Nvidia-optimized software
- Large memory for complex AI models
- Professional support through Nvidia forum
Risks
- High price compared to AMD competition
- Software not yet mature, desktop occasionally freezes
- Limited graphics output (only one display)
- Loud under load (2.2 sone)
Future Outlook
Short-term (1 year): Nvidia will fix software issues and deliver performance optimizations. More manufacturer variants will come to market.
Medium-term (5 years): Local AI development becomes standard, larger AI models require more RAM capacity. Stacking functions will be used more frequently.
Long-term (10-20 years): AI hardware becomes commodity, specialized mini PCs displace cloud services for sensitive AI applications. ARM architecture establishes itself in workstations.
Fact Check
Well documented: Technical specifications, benchmark results, price comparisons with concrete competing products are verified through practical tests.
Transparency lacking: Nvidia-internal performance claims use different test conditions. Long-term stability of the software is not yet evaluable.
Brief Verdict
The DGX Spark is a specialized AI developer machine, not intended for gaming or general workstation tasks. The device enables local development with large AI models in compact form for the first time, but only justifies the high price for CUDA-dependent projects. Nvidia still needs to work on software stability before the system is production-ready.
Three Key Questions
Freedom: How independent does local AI development make us from cloud providers and their data policies?
Responsibility: What ethical obligations arise when companies can develop their own AI models without external control?
Innovation: Can proprietary CUDA technology prevail long-term against open standards like ROCm, or does it hinder competition?