Magazine #3 | Autumn 2023
Water Is the New CO₂
With advances in Deep Learning, which uses artificial neural networks to analyze large data sets, Artificial Intelligence has become a real game changer. Advances in technological capabilities are leading to scientific breakthroughs and accelerating business growth, and they appear to be offering solutions to global challenges in important areas such as the climate crisis.
But the success of AI relies heavily on computationally intensive calculations to learn useful patterns from data during training and to check whether the predictions based on them are accurate during inference, the application phase of AI systems. As such, AI models, especially large generative models like GPT-3 and LaMDA, are typically trained on large clusters of servers, each with multiple graphic processing units (GPUs), creating an enormous appetite for energy. To curb the tremendous energy demand of AI systems, it is time we address their environmental footprints in their entirety.
While a low carbon footprint has now entered the public consciousness as an indicator of sustainability, the water footprint of AI systems – the fresh water consumed for generating electricity and cooling servers – is still given too little attention. Even putting aside the significant water toll of chip manufacturing, training a large language model like GPT-3 and LaMDA can easily evaporate millions of liters of fresh water for cooling the power plants and AI servers. This is all the more concerning as water becomes increasingly scarce due to rapid population growth and/or outdated water infrastructures, especially in drought-prone areas. Water scarcity has become one of our greatest global challenges. Despite all the efficiency gains being made in the field of Artificial Intelligence in terms of resource consumption, the exponential growth in demand is resulting in an ever-increasing water footprint. For example, Google’s direct water consumption increased by 20 percent between 2021 and 2022, and even doubled in certain drought-hit areas. Microsoft saw a 34-percent increase in its direct water consumption over the same period.
The Scope-2 indirect carbon footprint caused by electricity usage for training is routinely recorded in the model cards of published AI models. Yet, not even the direct water consumption during AI model training is included in the model card, not to mention the indirect water consumption tied to its electricity usage. To some extent, withholding information about AI’s water footprint is comparable to not including calorie content in the nutrition facts label of a food product. Such a lack of transparency is more than just an impediment to innovations that could improve water sustainability. It’s also difficult to reconcile with recent statements made by major technology companies in regard to water. Google, for example, announced its intention to become water neutral by 2030.
Developers of AI models need to take urgent action to curb growing water consumption. A first and crucial step would be to increase transparency and publicly disclose how much water is used to train and infer AI models, both directly for cooling AI servers and indirectly for generating the electricity to power them. An AI model’s water footprint should be recorded in its model card. Only with the availability of this information will it be possible to holistically benchmark AI’s environmental footprint. This measure would complement current efforts to make water supplies more sustainable, such as integrating more water-efficient techniques for cooling servers into AI systems. In addition, if the AI water footprint became more transparent, developers could exploit the spatial and temporal flexibilities of AI and train and deploy AI models in places where their footprint will be smaller. It also enables flexible trade-offs: If the AI model is deployed in a water-stressed area, it would probably make more sense to use a compact model with a smaller water footprint than a full, more resource-intensive model. Knowing AI’s water footprint data could also mitigate the environmental inequity that is accelerated by AI systems. We could move AI workloads around to equitably balance AI’s water footprint across different regions rather than letting a few disadvantaged and drought-stricken areas disproportionately bear the negative impact.
We can no longer allow the water footprint of AI systems to remain under the radar. It must be prioritized as part of the global fight against water scarcity. The first step is simple: We need to measure AI’s water footprint and make that information public.
Thirsty AI
ChatGPT needs about 500ml water for a simple conversation of 20-50 questions and answers. Since the chatbot has more than 100 million active users, each of whom engages in multiple conversations, ChatGPT’s water consumption is staggering. And it’s not only the application’s operational mode: Training GPT-3 in Microsoft’s state-of-the-art U.S. data centers would directly consume 700,000 liters of clean freshwater (enough for producing 370 BMW cars or 320 Tesla electric vehicles) and water consumption would be three times that if training were performed in Microsoft’s Asian data centers.
These estimates are taken from the yet to be peer-reviewed study Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. The study’s authors provide a methodology to evaluate the water footprint of AI models, using public data sources. They also explain how developers could reduce the water footprint of their AI models and increase water efficiency by scheduling AI model training and inference in different places and at different times.
SHAOLEI REN
Associate Professor of Electrical and Computer Engineering at the University of California, Riverside
His research broadly focuses on AI and Sustainability, with the goal of building an environmentally sustainable and equitable future.