The Reason Researchers Are Suddenly Focusing On AI Energy Use

In a socio-economic system suddenly constrained by historic financial disruptions and concerns about the impact of technology on respiratory health, understanding the true cost of AI may lead to more efficient resource allocation decisions, performance improvements and cost-benefit analyses. JL

Sara Castellanos reports in the Wall Street Journal:

Developing a single AI model can have a carbon footprint equivalent to the lifetime emissions of five average U.S. cars. The researchers’ goal is mitigate their carbon footprint by training algorithms on servers that run on hydroelectricity or solar power, using less training data overall and using pre-trained models, among other techniques. “If people see the true cost of these systems, we’d have a lot of harder questions about whether that convenience [of AI-based digital assistants, for example] is worth the cost.”

Developing artificial intelligence can use a lot of energy.

Researchers in Canada and the U.S. are developing tools to calculate the carbon footprint of AI models in an effort to help those who create them understand the impact the models are having on the environment.

Training AI models to understand human language, for example, requires data center servers and computer chips to process vast amounts of data and perform compute cycles and experiments that can run over days or weeks.

Developing a single AI model can have a carbon footprint equivalent to the lifetime emissions of five average U.S. cars, according to a paper published last year by researchers at the University of Massachusetts, Amherst. That is assuming the model runs in a data center powered by about 60% fossil fuels, combined with renewable and nuclear energy, said Emma Strubell, a lead author on the paper.

The new tools can give software developers and machine learning researchers an understanding of their models’ carbon emissions.

The researchers’ goal is to encourage their colleagues, data scientists and AI-focused software developers to mitigate their carbon footprint by training algorithms on servers that run on hydroelectricity or solar power, using less training data overall and using pre-trained models, among other techniques.

“If people could see the true cost of these systems, I think we’d have a lot of harder questions about whether that convenience [of AI-based digital assistants, for example] is worth the planetary cost,” said Kate Crawford, co-founder of the AI Now Institute at New York University, which focuses on examining the social implications of AI. Ms. Crawford, also a senior principal researcher at Microsoft Corp. ’s Research division, wasn’t involved in the development of the tools.

Sasha Luccioni, a postdoctoral researcher at Mila - Quebec AI Institute, is working with other researchers to develop a tool that lets users type a software command in their code that automatically generates an estimate of the amount of carbon dioxide emitted during training of an AI model. The tool, expected to launch in September, makes its calculation based on the specific energy grid the user is connected to, the type of processors that are used and the amount of time it takes to run the AI model, she said.

The carbon footprint estimates are also expected to be integrated into a dashboard made by startup Comet.ml, which provides data scientists with a platform to track, compare and optimize machine learning experiments and models. Data scientists can use the dashboard to compare the performance of various machine learning experiments as well as their energy consumption, she said.

Ms. Luccioni also worked with experts from AI software provider Element AI Inc. and the University of Montreal last year to develop an online carbon emissions calculator for machine learning researchers at universities and private companies. The calculator lets users input their own values such as hardware type, hours and cloud provider, to generate an estimate of carbon emissions usage, instead of automatically calculating emissions usage while the AI model trains. The researchers don't track usage of the tool, which is available online free of charge.

A team that includes researchers from Stanford University in California, Facebook Inc.’s AI Research division and McGill University in Montreal developed a similar calculator tool to measure the carbon footprint of machine learning experiments. Their “experiment impact tracker” tool was launched in February on Microsoft-owned GitHub, a website where software developers can collaborate on code.

The group aims to make improvements to the tool by the end of the summer, including adding tutorials on how to use the tool and expanding the number of hardware devices the tool can calculate carbon emissions from. Currently, the tool can calculate emissions from some processing units from Nvidia Corp. and Intel Corp.

“Our main goal was basically to raise awareness,” said Peter Henderson, a Ph.D. computer science student at Stanford who worked on the project. The team doesn’t track usage of the free tool, but Mr. Henderson said it might encourage researchers to mitigate their carbon footprint by choosing to run their training jobs on a clean energy grid.

Major cloud service providers offer options that allow users to run AI models on different energy grids, he said. Electricity in parts of Canada runs on hydroelectric power, while areas of the U.S. vary between solar power, coal, nuclear and other forms of power, according to electricityMap, an online tool made by Denmark-based climate action technology company Tmrow ApS.

The calculator tool could also encourage researchers to add carbon impact statements to their papers, Mr. Henderson said. That could make it easier to measure the energy usage consumed by AI models that are being presented at academic AI conferences.