The Language-Energy Divide: Measuring Energy Costs of Multilingual LLM Inference

This study introduces the ML.Energy framework to analyze the energy consumption of multilingual large language models (LLMs) during inference. It reveals that energy costs can vary significantly across languages, with disparities of up to 8.3 times per output token and up to 179 times for total energy consumption for a fixed set of requests, highlighting a systemic energy inequity in multilingual deployments. The findings underscore the importance of incorporating energy efficiency as a critical evaluation metric in LLM development and deployment, particularly for low-resource languages that exhibit both high energy costs and lower task accuracy.