Towards Green AI: Energy-Efficient Training and Inference of Large Language Models

Abstract:Large Language Models (LLMs) have transformed natural language processing with theirremarkable capabilities, yet their ever-growing computational demands come with steepenergy costs and increasing carbon emissions. This paper provides an in-depth,comparative case study titled “Towards Green AI: Energy-Efficient Training andInference of LLMs,” examining how efficiency-oriented methods can make large-scaleAI more sustainable. Using a qualitative, analytical approach, it draws on research papers,benchmarks, and technical reports published between 2020 and 2025 to highlight keystrategies for improving energy efficiency—such as low-precision computation(FP16/FP8), parameter-efficient fine-tuning (LoRA), low-bit quantized inference(ATOM), optimized attention kernels (FlashAttention-2), and inference accelerations likespeculative decoding. The study discusses the practical trade-offs of these techniques,outlines best practices for measuring efficiency, and explores their broader implicationsfor developing and deploying greener AI systems at scale. Keywords: Green AI, case study analysis, energy efficiency, large language models,quantization, FP8, LoRA