GraphAI CEO and KAIST School of Computing Professor Minsoo Kim’s research team has developed ‘GFlux,’ a framework capable of processing ultra-large-scale graph computations with 1 trillion edges at ultra-high speed using just a single GPU computer. While existing methods required connecting 25 computers and took approximately 2,000 seconds for high-difficulty computations, GFlux completed the same task in just 1,184 seconds in a single GPU environment, demonstrating world-class performance.
This was achieved by introducing ‘GTask’ scheduling, which optimizes computation units to overcome the GPU’s limited memory, along with integrated main memory-GPU memory management technology. The research also utilized a proprietary compression format called HGF to reduce data size by nearly half, and for the first time in a GPU environment, applied a 3-byte address system to reduce memory usage by 25%.
These technological capabilities are expected to become core driving forces in next-generation AI and data analysis fields that handle massive data — including Graph Retrieval-Augmented Generation (RAG), knowledge graphs, and vector indexing. GraphAI engineers Seyeon Oh and Heeyong Yoon participated as co-first authors, CTO Donghyung Han as the third author, and CEO Minsoo Kim as the corresponding author, demonstrating an exemplary case of industry-academia collaboration.
GraphAI plans to leverage the technology validated through this research to take the lead in solving real-world industrial challenges that require large-scale graph computation processing.
Read more : https://www.aitimes.kr/news/articleView.html?idxno=35071