The water footprint of AI models

A simple conversation of around 20-50 questions and answers using ChatGPT, AI model would need to "drink" a 500ml of water.

While a 500ml bottle of water might not seem like much, the total combined water footprint for inference becomes significant when considering billions of users.?

This whole article is written based on the paper Making AI Less thirsty other links provided in reference.

What is the water footprint of AI models?

The water footprint of AI models refers to the total amount of water consumed during their training and inference processes.

Why should we be concerned about the water footprint of AI models in data centers?

We should be concerned about the water footprint of AI models in data centers because, with the increasing demand for AI applications, the water consumption of data centers may continue to rise. This can exacerbate water scarcity issues and put additional stress on water resources, which are already under pressure from population growth, urbanization, and climate change.

Is it feasible to utilize sea or saltwater for data center cooling and electricity generation to reduce freshwater consumption?

Utilizing sea or saltwater for data center cooling and electricity generation is technically feasible, but there are challenges associated with corrosion, scaling, and biofouling that need to be addressed. Desalination processes can also be energy-intensive and have environmental implications. However, research and development in this area can potentially lead to more sustainable solutions.

How much water is currently being consumed during the training and inference processes of large language models (LLMs)?

The water consumption of large language models (LLMs) during training and inference processes varies depending on the specific models, data centers, and energy consumption. Estimates suggest that millions of liters of clean freshwater can be consumed by large AI models like GPT-3 and LaMDA.

Inferencing : A simple conversation of around 20-50 questions and answers using ChatGPT, AI model would need to "drink" a 500ml of water.

The water consumption of different AI models training, including GPT-3 (US), GPT-3 (Asia), and LaMDA:

• For GPT-3 (US) training, on-site water consumption is 700,000 liters, off-site water consumption is 2,800,000 liters, and the total water consumption is 3,500,000 liters.
• For GPT-3 (Asia) training, on-site water consumption is 2,100,000 liters, off-site water consumption is 2,800,000 liters, and the total water consumption is 4,900,000 liters.
• For LaMDA training, the water consumption varies and is in the range of millions of liters for both on-site and off-site consumption, resulting in a varying total water consumption.On-site water consumption refers to water used directly in data centers for cooling, while off-site water consumption refers to water used indirectly by data centers for electricity generation in power plants.

What is the formula for calculating the water footprint of AI models, and how was it derived?

The formula for calculating the water footprint of AI models is: Total Water Consumption = e [PUE EWIF + WUE_on], where e is the energy consumption, PUE is the power usage effectiveness, EWIF is the electricity water intensity factor, and WUE_on is the on-site water usage effectiveness. This formula is derived by considering various factors such as energy consumption, data center efficiency, and the amount of water used per unit of energy generated by power plants. Refer the paper given below in reference section for more details.

What are the primary data sources and methods used to estimate the water consumption of AI models, and how accurate are these estimates?

The primary data sources for estimating AI model water consumption include weather data, energy fuel mix data, water usage effectiveness values, and water consumption data from industries like car manufacturing. These estimates are approximate and depend on specific model and data center details.

Is the paper's argument about the water footprint of AI models convincing and well-supported?

The paper presents a well-researched argument supported by data. It emphasizes the importance of considering water footprint alongside carbon footprint in AI models and calls for transparency in reporting water efficiency.

The paper proposes a methodology for estimating AI models' water footprint and demonstrates their significant impact, highlighting the spatial-temporal diversity of water usage effectiveness (WUE) and suggesting that adjusting "when" and "where" AI models are trained can help reduce their water footprint. While the paper presents a strong case, its conclusions could be subject to further scrutiny and debate within the scientific community.

What proactive steps can be taken to minimize the water footprint of AI models and promote sustainable practices in the field?

And

What should applied data scientist know and do while using LLMs?

• Proactive steps to minimize the water footprint of AI models and promote sustainable practices in the field include:
• Improving data center efficiency and cooling technologies
• Using smaller, more efficient models when possible
• Fine-tuning pre-trained models instead of training from scratch
• Optimizing model usage and scheduling tasks efficiently
• Monitoring and optimizing resource usage
• Staying informed about sustainable AI practices and research
• Encouraging responsible AI development and usage within organizations and the broader AI community
• Collaborating with stakeholders to share insights and best practices
• Considering alternative approaches, such as federated learning or distributed learning, to reduce environmental impact.

AI and Water Use: Striking a Balance

I recently brought up the issue of AI models guzzling down water like there's no tomorrow, and someone in the discussion chimed in, saying that humans would drink 20 liters of water to do the same job, so it's all good. But wait a minute! Mother Nature has a VIP list for her water distribution, and it's reserved for living beings like humans, animals, and birds, not thirsty machines. We need to be responsible H2O stewards!

Imagine planting a data center underwater, thinking it solves everything. Guess what? It could still crank up the heat, turning the ocean into a giant jacuzzi. That's bad news for our fishy friends down there. So, let's not be too quick to give AI a free pass to the all-you-can-drink water buffet. Conservation is key!

-LLM assisted article

Reference :

1. Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models
2. A.I. tools fueled a 34% spike in Microsoft’s water consumption, and one city with its data centers is concerned about the effect on residential supply
3. AI water footprint suggests that large language models are thirsty
4. Artificial intelligence tech behind ChatGPT guzzles Iowa water, researchers say
5. Artificial Intelligence a guzzler as Microsoft's water use rises 34%
6. Does ChatGPT ‘drink’ water? Research reveals hidden cost of every conversation with AI chatbot