Large AIs can have a significant environmental impact because they rely on thousands of power-hungry computing servers housed within huge data centres. But the environmental damage could be reduced by better distributing the demands to different locations.
Such scheduling algorithms might lighten the AI workload on data centres in Arizona during summer droughts to reduce water-based cooling. It could also, in theory, favour a data centre region in Finland where , while minimising reliance on a Singapore data centre that has just 4 per cent carbon-free energy.
鈥淲hen data centres run their workloads, they are very flexible,鈥� says at the University of California, Riverside. 鈥淪o they can move the workloads around, which means shifting the environmental impact from one region to another.鈥�
Advertisement
Ren and his colleagues developed and tested a scheduling algorithm designed to minimise AI鈥檚 environmental impact on the most disadvantaged regions. Their algorithm balances energy cost, carbon footprint and water footprint associated with AI workloads when deciding how to distribute the workload across data centres 鈥� making use of regional advantages in carbon efficiency and water availability.
They also tested the algorithm against competing algorithms 鈥� focused only on reducing cost or minimising overall environmental impacts without accounting for regional differences 鈥� in a simulation that depicts 10 data centres spread across the US, Europe and Asia. The results showed that the main algorithm performed the best in achieving the lowest carbon and water footprints for the most disadvantaged regions 鈥� whereas competing algorithms that only minimised overall energy cost, carbon footprint or water footprint could sometimes worsen environmental inequity.
Such 鈥渟mart workload distribution strategies鈥� have 鈥済reat potential鈥� for reducing AI environmental impacts, especially if they take data centres鈥� on-site renewable power and AI performance flexibility into account, says at Villanova University in Pennsylvania. He also emphasises the importance of accounting for how regional water availability is impacted by both data centres鈥� on-site water needs and additional water consumption associated with the power plants that supply the data centres.
The strategy is theoretically feasible but would be difficult to coordinate across many different cloud computing providers, says at Hugging Face, a company that develops tools for sharing AI code and data sets. 鈥淢aybe within a single cloud provider it would be possible, but not across multiple ones,鈥� she says.
For long-term success, the strategy also requires more data and transparency from cloud providers before it can become practical, says Ren, along with greater political awareness around environmental equity so that data centre operators feel compelled to act.
Reference:
UC Riverside,
