LLM-Enabled Software Engineering Tools Advance Energy Efficiency Throughout the SDLC

Researchers are increasingly focused on the energy demands of artificial intelligence, but a critical area often overlooked is the power consumption of the software engineering tools used to create AI systems. Himon Thakur and Armin Moin, from the Department of Computer Science at the University of Colorado Colorado Springs (UCCS), alongside et al., investigate the energy efficiency of these AI-enhanced software development tools, specifically those leveraging large language models (LLMs). This work is significant because as AI features become standard in coding environments, energy use throughout the software development lifecycle could rise dramatically, this study proposes a novel approach combining Retrieval-Augmented Generation (RAG) with Prompt Engineering Techniques (PETs) to improve both code quality and energy efficiency, demonstrated through a comprehensive analysis of models ranging from 125M to 7B parameters.

RAG and Prompt Engineering for Efficient LLMs

Scientists have demonstrated a novel approach to significantly improve the energy efficiency of Large Language Models (LLMs) within Software Engineering (SE) tools, such as Computer-Aided SE (CASE) tools and Integrated Development Environments (IDEs). This research addresses a critical, yet largely unstudied, aspect of the increasing integration of AI into the software development lifecycle, the substantial energy consumption associated with these powerful AI capabilities. This work establishes a crucial step towards ‘ENERGY STAR’ LLM-enabled software engineering tools, minimising environmental impact without sacrificing performance.
The study reveals a comprehensive framework for measuring real-time energy consumption and inference time across a diverse range of LLM architectures, from 125M to 7B parameters. Models including GPT-2, CodeLlama, Qwen 2.5, and DeepSeek Coder were rigorously tested, providing a robust proof of concept and validating the core ideas behind the proposed approach. Researchers meticulously evaluated the impact of RAG pipelines on LLM performance, focusing on tasks like automated code snippet suggestions within modern IDEs. The experimental setup leveraged the CodeXGLUE dataset, specifically the CONCODE text-to-code subset, and Kaggle’s Natural Language to Python Code dataset, ensuring a diverse and representative evaluation of the framework’s capabilities.

Experiments show that the proposed RAG and PETs framework can substantially reduce the energy footprint of LLMs during code generation. Building upon prior work demonstrating the power of strategic prompt engineering, this research extends those findings by integrating a retrieval mechanism to enhance both efficiency and code quality. The framework dynamically selects the number of relevant code examples to include in the prompt, ensuring optimal performance within the LLM’s context window. Energy monitoring was conducted using the CodeCarbon library, providing precise measurements of energy consumption during inference, although the team rightly acknowledges the difficulty in accurately determining the source of the energy used.
This breakthrough directly addresses four key research questions: whether RAG can reduce energy consumption or inference time, how different LLM architectures compare in terms of energy efficiency, if a correlation exists between model size and RAG benefits, and whether RAG can enable smaller LLMs to achieve performance comparable to larger models while maintaining low energy usage. The work opens exciting possibilities for developing more sustainable and environmentally friendly software development tools. By optimising LLM performance and reducing energy demands, this research paves the way for a future where AI-powered SE tools are both powerful and responsible, contributing to a greener and more efficient software development ecosystem.

RAG and Prompt Engineering for Efficient AI-SE

Scientists investigated the energy efficiency of Software (SE) for AI-enabled Systems, focusing on Computer-Aided SE (CASE) tools and Integrated Development Environments (IDEs). This methodology addresses the growing concern that increasingly active, default-enabled AI features within development tools are significantly altering energy consumption patterns throughout the Software Development Lifecycle (SDLC). . Experiments employed a comprehensive framework designed to measure real-time energy consumption and inference time across a diverse range of Large Language Model (LLM) architectures.

The study meticulously profiled models varying in size from 125M to 7B parameters, specifically including GPT-2, CodeLlama, Qwen 2.5, and DeepSeek Coder, selected to provide sufficient validation and a proof of concept for future, more extensive analyses. Energy usage was quantified during code generation tasks, allowing for direct. The research team developed a framework to measure real-time energy consumption and inference time across LLMs ranging from 125M to 7B parameters, specifically GPT-2, CodeLlama, Qwen 2.5, and DeepSeek Coder. Experiments revealed that well-designed prompts, as previously reported by Rubei et al., can indeed reduce LLM energy consumption, alongside potential reductions in inference time.

Results demonstrate that CodeLlama achieved the most promising outcomes, exhibiting a 25% faster inference time and substantial improvements in code quality when utilising Retrieval-Augmented Generation (RAG). Conversely, DeepSeek Coder and Qwen showed increased energy consumption with RAG, although they generally produced higher quality code, presenting a clear trade-off between efficiency and performance. GPT-2 displayed slightly better energy efficiency but was slower in inference, highlighting the varied impacts of RAG across different model architectures. The team meticulously measured energy consumption levels, finding GPT-2 to be the most efficient, followed by CodeLlama, while DeepSeek Coder and Qwen consumed approximately three times more energy.

Data shows no clear correlation between model size and RAG-based energy efficiency benefits; only GPT-2 and CodeLlama demonstrated energy reduction with RAG, irrespective of the other models’ size. Specifically, GPT-2, with RAG on the Kaggle dataset, achieved a code quality score of 0.6, matching DeepSeek Coder’s performance while consuming approximately 3.5times less energy. This breakthrough delivers a significant finding: RAG can enable smaller, more efficient models to achieve competitive code generation quality. Furthermore, tests prove that Qwen, without RAG, was the fastest for inference, followed by GPT-2, DeepSeek Coder, and CodeLlama. The study’s outcomes confirm that RAG’s impact varies considerably across LLMs, offering valuable insights for optimising energy usage in SE tools and paving the way for more sustainable AI-driven software development practices. Future work will explore the use of more powerful servers and incorporate established code quality metrics like CodeBleu, alongside static and dynamic analyses, to further refine code quality assessments.

RAG Impacts LLM Energy and Speed

Scientists have investigated the energy efficiency of AI-enhanced software engineering tools, focusing on the impact of Large Language Models (LLMs) on energy consumption during code generation. A comprehensive framework was developed to measure real-time energy consumption and inference time across various LLM architectures, including GPT-2, CodeLlama, Qwen 2.5, and DeepSeek Coder, ranging in size from 125M to 7B parameters. The findings demonstrate that the effects of RAG varied significantly between the tested LLMs.

CodeLlama experienced a 25% reduction in inference time alongside improvements in code quality, while smaller models like GPT-2 showed mixed results, achieving modest energy savings but not consistently improved performance. Notably, GPT-2, when paired with RAG, matched the code quality of the larger DeepSeek Coder model while consuming approximately 3.5times less energy, suggesting that RAG can enable smaller, more efficient models to achieve competitive results. The authors acknowledge limitations related to the server infrastructure used for measurements, highlighting the need for further investigation using cloud servers to account for potential external influences. Future research will incorporate established code quality metrics, such as CodeBleu, alongside static and dynamic analyses to provide a more thorough assessment of generated code. The team also intends to explore combining RAG with the Model Context Protocol (MCP) and investigate the energy efficiency implications for quantum computing, specifically Python code generation for quantum SDKs. This work contributes to a growing understanding of the energy footprint of AI-powered software development tools and suggests that strategic use of techniques like RAG can help mitigate energy consumption without sacrificing code quality.