[Paper Review] Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models

Pengfei Li, Jianyi Yang+2 more|arXiv (Cornell University)|Apr 6, 2023

Water-Energy-Food Nexus Studies148 citations

TL;DR

This paper presents a principled methodology to estimate the total water footprint of AI models, including operational (scope-1/2) and embodied water, and demonstrates how water efficiency varies spatially and temporally to inform scheduling strategies. It advocates transparency and holistic consideration of water and carbon footprints for sustainable AI.

ABSTRACT

The growing carbon footprint of artificial intelligence (AI) has been undergoing public scrutiny. Nonetheless, the equally important water (withdrawal and consumption) footprint of AI has largely remained under the radar. For example, training the GPT-3 language model in Microsoft's state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret. More critically, the global AI demand is projected to account for 4.2-6.6 billion cubic meters of water withdrawal in 2027, which is more than the total annual water withdrawal of 4-6 Denmark or half of the United Kingdom. This is concerning, as freshwater scarcity has become one of the most pressing challenges. To respond to the global water challenges, AI can, and also must, take social responsibility and lead by example by addressing its own water footprint. In this paper, we provide a principled methodology to estimate the water footprint of AI, and also discuss the unique spatial-temporal diversities of AI's runtime water efficiency. Finally, we highlight the necessity of holistically addressing water footprint along with carbon footprint to enable truly sustainable AI.

Motivation & Objective

• Motivate the need to study the hidden water footprint of AI models amid freshwater scarcity.
• Develop a principled methodology to estimate both operational and embodied water footprints of AI models.
• Showcase a case study estimating GPT-3’s operational water consumption to illustrate methodology.
• Highlight the spatial-temporal variability of water efficiency and its implications for scheduling AI workloads.
• Advocate for transparency and holistic sustainability, integrating water and carbon footprints.

Proposed method

• Define and distinguish water withdrawal vs. water consumption (WWF vs. WCF).
• Model operational water footprint using on-site WUE (scope-1) and off-site WUE (scope-2) with time-varying factors.
• Incorporate data center PUE to relate IT energy use to total water use.
• Compute embodied water footprint as amortized manufacturing water over server lifespan.
• Combine operational and embodied components to obtain total water footprint (WaterTotal).
• Apply the framework to a GPT-3 case study using location-specific PUE/WUE and electricity water intensity data.

Experimental results

Research questions

• RQ1 How can we quantify the total water footprint (operational plus embodied) of AI models?
• RQ2 How do on-site and off-site water usage efficiencies vary across time and location, and how does this affect AI water footprint?
• RQ3 What are the implications of water footprint variability for scheduling training and inference workloads?
• RQ4 What transparency measures are needed to communicate AI water footprints to developers and users?
• RQ5 How should water footprint considerations be balanced with carbon footprint objectives in sustainable AI?

Key findings

• Operational water footprint for GPT-3 can be substantial and varies by location and cooling configuration (e.g., on-site and off-site water usage impacts).
• Water efficiency is spatially and temporally diverse, influencing optimal timing and location for training and inference to reduce water footprint.
• A holistic view shows potential conflicts between minimizing water footprint and carbon footprint, necessitating balanced strategies.
• Embodied water in manufacturing contributes to total footprint and is amortized over server lifespan in the model’s total life cycle.
• Transparency gaps exist in model cards; the authors advocate including scope-1 and scope-2 water usage information to improve understanding and stewardship.