|
Abstracto :
|
This supplement provides five examples of environmental impact assessments of AI systems, each addressing a distinct stage of the AI lifecycle management (AI-LCM). The examples show that data preprocessing, training, inference, monitoring & adaptation and data/model transmission can each constitute a significant source of energy consumption. The first four stages have been assessed through measurement. Transmission is assessed analytically based on measured model sizes and published network energy intensities. The findings motivate a comprehensive lifecycle energy metric that extends beyond the training and inference stages commonly considered in AI energy assessments. The proposed methodology applies the approach specified in ITU-T L.1801, to a large extent based on measured electricity consumption. By accounting for all relevant stages of the AI lifecycle, this approach enables a more comprehensive and fair assessment and comparison of the energy consumption of different AI systems.
|
Word
