Scope
In (Patterson et al, 2021) focuses on the environmental impact of large-scale neural network training, especially in terms of energy consumption and carbon emissions. With the exponential growth of computational demands in training AI models, particularly neural networks, the energy required has increased dramatically. As AI/ML systems evolve, the environmental footprint associated with them is becoming a significant concern. The authors explore the direct relationship between neural network training, energy use, and carbon emissions. They suggest that while the energy demand continues to rise, improvements in efficiency, hardware, and methodologies might eventually curb the emissions trajectory.
Summary
At first, this study notes that shifting AI training to times and locations with lower carbon intensity can significantly reduce emissions. New hardware (e.g., energy-efficient GPUs) and optimised software can help reduce the energy required for neural network training. As larger AI models require exponentially more energy, so balancing model size and performance with energy efficiency is critical for sustainability.
Relevance for EXIGENCE
This work is highly relevant to EXIGENCE project for energy metrics, energy measurements , energy optimisation, green incentives and incentive mechanism; aimed at incentivising users to consume less energy and reduce carbon emissions in their use of internet resources. In particular, users can be incentivised to schedule AI training during periods of low-carbon intensity or in regions where renewable energy is prevalent, reducing the environmental impact of their activities. Further, encouraging users to adopt lightweight, energy-efficient models (e.g., MobileNet over ResNet) could be part of a broader incentive structure that rewards sustainable computing practices. Additionally, by integrating real-time carbon footprint tracking, users can be made aware of their energy consumption and emissions, providing an incentive to make more environmentally friendly choices.
- Patterson et al., “Carbon Emissions and Large Neural Network Training,” arXiv:2104.10350 [cs], Apr. 2021.
