As the computing demands for artificial intelligence, or AI, grow at a great pace, so do the inequitable environmental consequences.
The rising computer processing demands from AI are increasing freshwater consumption to cool thousands of servers housed in warehouse-sized data centers and increasing unhealthful air pollution from coal power plants that provide much of the electricity.
A recent paper by University of California, Riverside, electrical and computer engineers finds that technology companies are not doing enough to equitably distribute these growing environmental impacts. The finding mirrors calls from international organizations such as the United Nations Educational, Scientific and Cultural Organization and the Organization for Economic Cooperation and Development for efforts to address AI’s environmental inequity.
The paper offers models and potential solutions to how these companies, such as Google and Microsoft, can distribute the processing loads for more equitable impacts across the globe.
Unlike many other industries, big tech has the flexibility to avoid regional environmental injustices, said UCR Bourns College of Engineering associate professor Shaolei Ren, the corresponding author of “Towards Environmentally Equitable AI via Geographical Load Balancing,” which has been released by the University of California as a preprint.
“They have freedom,” Ren said. “They can route workloads to different locations, and they can do that immediately.”
Instead, tech companies often use their flexibility to find the lowest capital and operating costs.
“It just does not feel right,” Ren said. “We are exploiting the cheaper prices and amplifying the environmental impacts.”
For example, several tech companies are planning to build large scale data centers in the Phoenix, Ariz. Region. Each is expected to use as much as 1 million to 5 million gallons of water a day — even though this desert region is increasingly strapped for water because of declining flows of the Colorado River.
“They don’t have to run lots of disproportionately power-hungry servers in Arizona to exhaust their water resources,” he said. “They can train the AI models in other places. They can route their workloads from Arizona to, let’s say, Washington.”
Beyond the use of fresh water, the miles of racks of servers in warehouse-size data centers consume increasing amounts of electricity produced at power plants that emit not only earth-warming carbon but also other air pollutants, such as particulate matter and nitrogen oxides, which react with other pollutants to form lung-irritating ozone. Such air pollutants have been linked to higher risks for cancer, heart disease, shorter lives, and other ill health effects.
A result is that people living near the power plants suffer the brunt of these consequences.
UCR doctoral candidates Pengfei Li and Jianyi Yang, are the first and second authors of the paper. Cal Tech professor Adam Wierman is co-author and Ren is the corresponding author. It is available as a preprint at eScholarship Publishing.
“AI will have some negative environmental impacts,” Ren said. “But, what’s more concerning for me is that the environmental cost is not equitably distributed across all data centers.
"We should reduce the environmental inequity to make AI truly responsible.”
Original source can be found here