A systematic energy profiling of on-device Vision-Language Model (VLM) inference across five models and two hardware platforms reveals that the primary energy bottleneck is output generation rather than visual perception. The study overturns the common assumption that reducing visual tokens is the most effective efficiency strategy, demonstrating instead that controlling output length is critical for power savings.
- Average inference power remains a model-intrinsic constant with less than 5% variation regardless of input resolution or image complexity.
- Each output token requires 11 to 39 times more wall-clock time than each input token due to compute and memory asymmetry between prefill and decode phases.
- Image complexity, measured by object count, causes up to 4.1x energy differences solely through variations in output length rather than visual processing cost.
- Visual token pruning saves at most 10% of total energy for fixed-token models, whereas controlling output length can save up to 97% across models ranging from 1 billion to 8 billion parameters.
These findings indicate that the true energy bottleneck in edge VLM inference is determined by how much the model generates rather than what it perceives, highlighting the need to prioritize decoding efficiency over visual token reduction.