[Paper Review] How Green Can AI Be? A Study of Trends in Machine Learning Environmental Impacts

Clément Morand, Anne‐Laure Ligozat+1 more|arXiv (Cornell University)|Dec 23, 2024

Air Quality Monitoring and Forecasting4 citations

TL;DR

This study analyzes the environmental impacts of training machine learning models and manufacturing graphics cards from 2013 to 2023, finding that despite efficiency improvements and strategies like location shifting to low-carbon regions, both hardware production and training energy use have increased exponentially. The rebound effect—where efficiency gains fuel larger models—undermines impact reduction, indicating that current strategies cannot curb AI's growing environmental footprint without reducing AI scale itself.

ABSTRACT

The compute requirements associated with training Artificial Intelligence (AI) models have increased exponentially over time. Optimisation strategies aim to reduce the energy consumption and environmental impacts associated with AI, possibly shifting impacts from the use phase to the manufacturing phase in the life-cycle of hardware. This paper investigates the evolution of individual graphics cards production impacts and of the environmental impacts associated with training Machine Learning (ML) models over time. We collect information on graphics cards used to train ML models and released between 2013 and 2023. We assess the environmental impacts associated with the production of each card to visualize the trends on the same period. Then, using information on notable AI systems from the Epoch AI dataset we assess the environmental impacts associated with training each system. The environmental impacts of graphics cards production have increased continuously. The energy consumption and environmental impacts associated with training models have increased exponentially, even when considering reduction strategies such as location shifting to places with less carbon intensive electricity mixes. These results suggest that current impact reduction strategies cannot curb the growth in the environmental impacts of AI. This is consistent with rebound effect, where the efficiency increases fuel the creation of even larger models thereby cancelling the potential impact reduction. Furthermore, these results highlight the importance of considering the impacts of hardware over the entire life-cycle rather than the sole usage phase in order to avoid impact shifting. The environmental impact of AI cannot be reduced without reducing AI activities as well as increasing efficiency.

Motivation & Objective

• To assess the evolving environmental impacts of graphics card production from 2013 to 2023.
• To analyze the environmental footprint of training ML models over the same period using real-world data.
• To evaluate the effectiveness of current impact reduction strategies such as hardware upgrades and location shifting.
• To investigate the role of rebound effects and impact shifting in undermining environmental gains in AI.

Proposed method

• Collected data on NVIDIA workstation graphics cards released between 2013 and 2023 to assess their production-phase environmental impacts.
• Used the Epoch AI dataset to link specific ML models to the hardware used for training and estimate training energy consumption.
• Applied MLCA (Machine Learning Carbon Assessment) to quantify carbon footprint, metallic resource depletion, and primary energy demand across the hardware life cycle.
• Modeled training energy use based on GPU count, training duration, and power consumption, adjusting for electricity mix and PUE (Power Usage Effectiveness).
• Evaluated the impact of location shifting to low-carbon electricity regions and found it insufficient to offset exponential growth.
• Accounted for uncertainties via sensitivity analysis, including assumptions about PUE, training duration, and electricity mix.

Experimental results

Research questions

• RQ1 How have the environmental impacts of graphics card production changed from 2013 to 2023?
• RQ2 To what extent do current efficiency strategies reduce the carbon footprint of ML model training?
• RQ3 Does the rebound effect negate the benefits of improved hardware efficiency in AI training?
• RQ4 How does impact shifting from use phase to manufacturing phase affect the overall environmental footprint of AI systems?
• RQ5 Can greener electricity mixes fully offset the growing energy demand from AI training?

Key findings

• The environmental impact of graphics card production increased continuously from 2013 to 2023, driven by rising complexity and material intensity.
• Despite improvements in GPU energy efficiency, the energy consumption and carbon footprint of training ML models have grown exponentially over the same period.
• Strategies like shifting training to regions with lower-carbon electricity mixes failed to curb the exponential rise in carbon emissions due to the rebound effect.
• The rebound effect is prevalent: increased efficiency leads to larger models and higher overall energy use, nullifying potential environmental gains.
• Impact shifting from use to manufacturing phase is significant, highlighting the need for full life-cycle assessments beyond just operational energy use.
• Current reduction strategies are insufficient to reverse the trend; reducing AI activity is necessary to meaningfully lower environmental impacts.