Vision-language models are moving from data centers to edge devices, but their energy footprint on small hardware remains largely unquantified. A new arXiv preprint takes up this question, focusing on single-board computers equipped with NPU or GPU accelerators.
The authors measure total-system energy during VLM inference, rather than only accelerator power. The abstract notes that little is known about how much energy these models draw on such platforms, or where that energy goes. The study appears to be an initial characterization, though the abstract does not yet report specific findings or numbers.
Because the abstract is truncated, this digest cannot detail results. Still, the work highlights a gap: as VLMs are embedded in robots, cameras, and other edge systems, knowing their energy profile is critical for battery life and thermal design.