Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models (2304.03271v5)

Abstract: The growing carbon footprint of 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 LLM 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.

Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models

The paper "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models" investigates an underexplored dimension of AI sustainability: the water footprint associated with AI model training and deployment. The paper outlines a comprehensive methodology to estimate both the operational and embodied water footprints of AI models, using GPT-3 as a case paper.

Overview

While the carbon footprint of large AI models has received substantial attention, the water footprint—defined as water withdrawal and consumption required for cooling and electricity generation in data centers—has largely remained unaddressed in academic and public discourse. The paper estimates that training a model like GPT-3 in Microsoft's state-of-the-art U.S. data centers can evaporate around 700,000 liters of water directly, with even more significant global implications as AI adoption expands. The authors predict that by 2027, the global AI demand could be responsible for 4.2 to 6.6 billion cubic meters of water withdrawal, potentially leading to serious environmental and societal impacts.

Methodological Approach

The authors present a dual-faceted methodological approach:

1. Operational Water Footprint: This includes on-site water used for server cooling (scope-1) and off-site water used for electricity generation (scope-2). By using variables such as Water Usage Effectiveness (WUE) and Energy Water Intensity Factor (EWIF), the paper provides an equation to compute the real-time water footprint of AI operations, taking into account spatial and temporal diversities.
2. Embodied Water Footprint: This relates to water used in manufacturing AI servers and is amortized over the lifespan of these servers. Although detailed scope-3 water usage data remains largely obscure, the authors emphasize its significance.

Key Findings

1. High Water Consumption: Training GPT-3 can consume up to 5.4 million liters of water, with operational scope-1 consumption reaching around 700,000 liters.
2. Significant Spatial and Temporal Variations: Water efficiencies vary considerably based on geography and time, influenced by local weather conditions and changing energy fuel mixes.
3. Transparency and Reporting: Current AI model cards lack comprehensive details about water usage, focusing almost exclusively on carbon footprints.
4. Potential Conflicts: Strategies to optimize for carbon efficiency may not necessarily align with those for water efficiency, sometimes even exacerbating water usage.

Practical and Theoretical Implications

Practical Implications

1. Dynamic Scheduling: Recognizing that both time and location matter, AI workloads should be dynamically scheduled to operate in water-efficient regions and times. This could involve training models during off-peak hours or in cooler climates to maximize water and energy efficiency.
2. Policy and Transparency: Increased reporting and transparency are crucial. Including water footprint data in AI model cards will help both developers and end-users make informed decisions, promoting better sustainability practices.
3. Infrastructure Design: Data centers should integrate water-efficient cooling solutions and explore alternatives like air-side economizers and purified non-potable water for cooling.

Theoretical Implications

1. Holistic Sustainability Models: The paper calls for holistic approaches that consider both carbon and water footprints to create genuinely sustainable AI solutions. This requires new frameworks that can balance these two metrics, potentially reconciling the trade-offs between them.
2. Innovation in AI Deployment: The paper encourages further research into optimizing AI's lifecycle—from manufacturing to deployment—minimizing water usage without compromising performance.

Speculation on Future Developments in AI

1. Data-Centric Transparency: Future AI ecosystems will likely see increased data transparency, driven by both regulatory requirements and consumer demand for sustainable technologies.
2. Advanced Eco-Friendly Architectures: New AI hardware could emerge that inherently designs for both carbon and water efficiency, using innovative cooling technologies and renewable energy to operate sustainably.
3. Synthetic Data Utilization: The advancements in AI techniques like Synthetic Data Generation could reduce the need for large-scale training on real data, subsequently lowering energy and water requirements.

Recommendations

• Schedule Awareness: Utilize time-based scheduling to train models during water-efficient periods.
• Model Card Updates: Extend AI model cards to include full environmental impact disclosures—covering both water and carbon footprints.
• Explore Hybrid Cooling Technologies: Adopt dynamic, hybrid cooling systems in data centers to better manage environmental conditions and reduce water usage.

Conclusion

This paper effectively uncovers the hidden water costs of AI development and calls for immediate action in addressing these issues. Its rigorous methodological framework for estimating water footprints and the practical and theoretical insights serve as valuable contributions toward making AI more socially responsible and environmentally sustainable. Addressing water usage as part of AI's environmental impact can no longer be an overlooked aspect; it is imperative for future AI advancements.

References

The list of references used in the paper provides rich context and validation for the claims made. This essay would be complemented by citing key studies from the reference list provided in the original paper.