DGX Spark hits 3–4秒 alerts, but 45～50W idle

The author had previously built a local-AI surveillance system on a mini-PC with a 32GB memory limit, using the lightweight Qwen 3-VL 4B Instruct model and relying on its internal GPU after finding an external GPU too bulky and power-hungry to run continuously. He then ported the same setup to the 128GB DGX Spark to see whether the extra headroom changed anything, switching to the larger Qwen 3-VL 30B A3B model and adding features such as real-time video commentary, half-day detection summaries, and a face-recognition filter that suppresses alerts for familiar people.

Comparing the two models on the Spark, he found almost no difference in the short push-notification text, which was capped at 160 tokens, but the 4B model could loop the same sentence when asked for longer descriptions. The 30B model's value, in his view, is mainly this greater stability. On speed, the Spark finished notification processing in 3～4 seconds, while the mini-PC took roughly 10 seconds with its internal GPU and about 5 seconds with the external GPU for the explanatory notification. The flip side is standby power: the Spark idles at 45～50W, two to three times the mini-PC's internal-GPU figure, with peaks around 100W.

The author's own verdict is that the Spark is more machine than a home camera system needs. The real trade-off hinges on whether a buyer values the faster, more capable analysis and additional features enough to accept the higher always-on power draw and the need for external storage for continuous recording. His suggested middle path is to run the surveillance workload on a NAS and use the DGX Spark instead as a development base for a wider smart-home control system built on Home Assistant.