energy-per-prompt-classification-vs-generation-vs-image
IN premise — entries/2026/06/21/wiki-Large_language_model-chunk-3.md
Created 2026-06-21T09:50:09+00:00
Energy cost per prompt varies by task type: classification ~0.002-0.007 Wh, text generation ~0.05 Wh, image generation ~2.91 Wh (up to 11.49 Wh for least efficient models).
Summary
The energy cost of a single prompt can swing by over two thousand times depending on what kind of output you're asking for, from a few thousandths of a watt-hour for a quick classification to nearly eleven watt-hours for an image. This means the single biggest lever for cutting a system's energy footprint is choosing the right task type, not fine-tuning model parameters or prompt length.
Dependents
These beliefs depend on this one:
- IN llm-resource-consumption-quantifiable-across-three-dimensions — LLM resource consumption is precisely quantifiable across three independent dimensions — memory (2 bytes per float16 parameter), compute (6 FLOPs per parameter per token for training, 1-2 for inference), and energy (0.002-0.05 Wh per prompt depending on task type) — enabling principled cost modeling for deployment decisions.