Green LLM Techniques in Action: How Effective Are Existing Techniques for Improving the Energy Efficiency of LLM-Based Applications in Industry?
Pelin Rabia Kuran, Rumbidzai Chitakunye, Vincenzo Stoico, Ilja Heitlager, Justus Bogner
TL;DR
This paper addresses the practical energy efficiency of LLM-based industrial applications by empirically evaluating four green techniques—Small/Large LM Collaboration, Prompt Optimization, Quantization, and Batching—applied to the ChatSBP chatbot within a controlled Leaplab environment. Using GSM8K and MMLU as benchmarks, the study conducts a full factorial experiment across 18 configurations, measuring energy consumption, accuracy, and response time. It finds that NPCC-based dynamic routing and 2-bit quantization can substantially reduce energy, but many techniques incur unacceptable accuracy or latency penalties, highlighting the trade-offs and the challenge of achieving energy-efficient LLM inference in real-world settings. The work provides actionable guidance for industry practitioners, demonstrating that while energy reductions are feasible, balancing them with quality attributes requires careful technique selection, hardware-aware experimentation, and further research on preserving accuracy alongside energy gains.
Abstract
The rapid adoption of large language models (LLMs) has raised concerns about their substantial energy consumption, especially when deployed at industry scale. While several techniques have been proposed to address this, limited empirical evidence exists regarding the effectiveness of applying them to LLM-based industry applications. To fill this gap, we analyzed a chatbot application in an industrial context at Schuberg Philis, a Dutch IT services company. We then selected four techniques, namely Small and Large Model Collaboration, Prompt Optimization, Quantization, and Batching, applied them to the application in eight variations, and then conducted experiments to study their impact on energy consumption, accuracy, and response time compared to the unoptimized baseline. Our results show that several techniques, such as Prompt Optimization and 2-bit Quantization, managed to reduce energy use significantly, sometimes by up to 90%. However, these techniques especially impacted accuracy negatively, to a degree that is not acceptable in practice. The only technique that achieved significant and strong energy reductions without harming the other qualities substantially was Small and Large Model Collaboration via Nvidia's Prompt Task and Complexity Classifier (NPCC) with prompt complexity thresholds. This highlights that reducing the energy consumption of LLM-based applications is not difficult in practice. However, improving their energy efficiency, i.e., reducing energy use without harming other qualities, remains challenging. Our study provides practical insights to move towards this goal.