Study finds LLM prefill compute time and energy scale quadratically with input length

[arXiv] A preprint paper by Krupp and co-authors says the authors systematically profiled compute time and energy consumption across a wide range of prefill (input) and decode (output) token lengths for large language models. Using generalized linear models, the authors report that decode compute time and energy scale linearly with output length, while prefill compute time and energy show both linear and quadratic scaling with input length, indicating that input context is a relevant factor for large-context workloads. The authors also say compute time serves as a reliable proxy for energy consumption when direct energy measurements are unavailable, and that their mapping framework enables precise energy estimation and supports more sustainable deployment decisions for agentic AI. The work is a preprint and has not been peer-reviewed.