Energy-Aware Prompt Engineering for Large Language Models: An Empirical Study on Software Engineering Tasks

The growing environmental impact of AI-based software systems, particularly those leveraging large language models (LLMs), necessitates urgent investigation into their substantial resource consumption, which contributes significantly to data center load and carbon emissions. This paper extends our prior work on LLM energy consumption, which focused solely on Llama 3.0 for code completion, by empirically analyzing the influence of Prompt Engineering Techniques (PETs) on the energy footprint of two prominent LLMs: the general-purpose Llama 3.1 and the specialized Code Llama. Our investigation spans two distinct natural language tasks: code completion and textual summarization of GitHub README.md files. Specifically, we examine the impact of custom prompts on code completion and the effect of input formatting (comparing plain text, Markdown, and HTML representations) on the textual summarization task. Our findings indicate that prompt customization positively influences the energy efficiency of both Llama 3.1 and Code Llama. However, for the textual summarization task, the specific input format (plain text, Markdown, HTML) appears to have a less pronounced contribution to energy consumption. Notably, Code Llama consistently exhibited a substantially higher energy footprint compared to Llama 3.1 throughout the majority of our experimental evaluations, highlighting potential differences in resource demands between general-purpose and specialized models.