Measurements and optimizations of the energy consumption of AI systems

The increasing energy consumption of Large Language Models (LLMs) poses a significant challenge to their sustainable deployment. Service operators face a fundamental conflict: maximizing inference throughput often leads to prohibitive increases in total energy consumption and job duration. This report presents a rigorous, multi-platform analysis to resolve this performance-versus-efficiency dilemma. We investigate the impact of Activation-aware Weight Quantization (AWQ), a stateof-the-art quantization technique, on the performance and energy consumption of modern LLMs. Our experiments are conducted across three distinct NVIDIA GPU architectures: Hopper H200, Ampere A100, and Ada RTX 5000. Our results confirm that 4-bit AWQ is a dominant optimization, delivering substantial, concurrent gains: throughput is increased by up to 118.5% while energy consumption is reduced by up to 54.4%, all without catastrophic degradation in model accuracy. Our primary contribution is a practical, two-step methodology for optimal deployment: first, apply AWQ quantization, and second, select the batch size that minimizes the Energy-Delay Product (EDP). This strategy allows service operators to achieve a quantifiably optimal balance between performance and total cost, providing a validated path towards the sustainable deployment of AI systems.