Korea University Says It Has Developed AI Training Task Scheduling Technology That Reduces GPU Waste
- Input
- 2026-08-25 15:56:50
- Updated
- 2026-08-25 15:56:50

[Financial News] A study has found that the GPU task management method long used to speed up AI training can actually slow training down.
A research team led by Professors Gyeongsik Yang and Hyuck Yoo in the Department of Computer Science at Korea University’s College of Informatics said on the 25th that it had identified a phenomenon called 'futile preemption' during GPU task processing and developed a task scheduling technology called 'Lazer' to reduce it.
Training large-scale AI models requires hundreds to thousands of GPUs. Companies and research institutes build clusters so multiple training jobs can share limited GPUs, and schedulers determine the order in which tasks run. Until now, the widely used approach has been 'preemption,' in which a running task is paused and its GPU is handed over when a higher-priority task arrives. But if that new task is itself pushed aside by yet another higher-priority job before it even begins execution, the time spent preparing it is wasted entirely.
The Korea University researchers reproduced six months of operating logs from a large-scale GPU cluster under various scheduling methods and confirmed the phenomenon. They named it 'futile preemption.'
AI training jobs require preparation time ranging from several minutes to tens of minutes before they actually begin running on GPUs. After analyzing the operating logs, the team found that time wasted by futile preemption accounted for as much as 13% of total GPU usage time. That is equivalent to leaving 512 GPUs unused for about 35 days.
Lazer, the system developed by the researchers, considers the interval between incoming tasks, the remaining runtime of existing tasks, and the time needed to prepare execution after a task is placed on a GPU. Based on these factors, it decides whether to replace a task or delay preemption.
Tests on an actual GPU server showed that futile preemption was eliminated, and the time from task submission to completion was reduced by an average of 3.8 times and by as much as 8.8 times compared with the existing method.
Professor Gyeongsik Yang said, "Contrary to the common belief that preemption speeds up training, we confirmed that it can instead cause GPU waste and delay learning." He added, "This study shows that AI infrastructure efficiency can be improved through system software improvements alone."
Meanwhile, the study was accepted for presentation at DAC (Design Automation Conference) 2026, held in Long Beach, the United States, from July 26 to 29. The paper was also selected as a candidate for the Best Paper Award, with 13 papers chosen from more than 2,400 submissions this year.
[email protected] Lee Bomi Reporter