A way to reduce energy consumption during AI training by 30% is proposed

Researchers have found a way to reduce “idleness” electricity consumption.

Researchers from the University of Michigan show that it is possible to reduce the energy required to train large AI language models such as GPT by 30%. The proposed method makes it possible to complete training in the same time, consuming less electricity.

Scientists explain that energy costs arise due to the uneven distribution of calculations between graphics processors. Some processors are overloaded while others are idle, resulting in inefficient power consumption.

AI models are so large that they cannot fit into a single computer processor. They need to be divided into tens of thousands of processors for training, but it is almost impossible to divide the models into perfectly equal sizes, – says Jae-Won Chung, co-author of the study

200% Deposit Bonus up to €3,000 180% First Deposit Bonus up to $20,000

Learning tasks are also difficult to distribute evenly because some tasks must be grouped together on the same processor. Depending on the grouping, some processors may get stuck on complex tasks, while others quickly solve simple ones and idle.

To solve this problem, the researchers developed the Perseus program. It identifies critical areas of calculations or a number of subtasks that will take the most time to complete. The program then slows down processors that are not on the critical path so that they all complete their work at roughly the same time, eliminating unnecessary energy consumption. do not change, and energy consumption is reduced by 30%. The saved resources will be enough to power 1.1 million homes. In addition, the proposed method will reduce the carbon footprint associated with the operation of data centers.