GraphAI CEO and KAIST School of Computing Professor Minsu Kim’s research team has developed ‘FlexGNN,’ a technology that enables ultra-fast training of ultra-large-scale GNN models on just a single GPU server, breaking away from conventional approaches that required multiple GPU servers. The system applies database query optimization techniques to AI training, optimally controlling data flow between GPU, main memory, and SSD, thereby resolving memory shortage issues and improving training speed by up to 95x compared to existing technologies.
This breakthrough enables full-graph AI implementation that surpasses supercomputers in fields requiring precise analysis, such as climate prediction and new material discovery. GraphAI researcher Jeongmin Bae participated as the first author, CTO Donghyung Han as the second author, and CEO Minsoo Kim served as the corresponding author.
The FlexGNN technology developed by the research team was presented at ‘ACM KDD,’ the world’s most prestigious data mining academic conference, and will be applied to GraphAI’s next-generation graph database solution ‘GraphOn’ to dramatically improve data analysis efficiency in real industrial settings.