SEOUL, May 29 (AJP) - Researchers at the Korea Advanced Institute of Science and Technology in South Korea have developed a virtual platform to test the performance of large artificial intelligence server setups without building physical systems, the prominent research institute said Friday.
Created by a team led by Professor Park Jong-se at the Korea Advanced Institute of Science and Technology (KAIST) School of Computing, the software is called LLMServingSim 2.0. It models the complex hardware and software interactions needed to run large language models. The tool allows developers to experiment with different designs and verify performance before buying expensive server equipment.
Running large artificial intelligence models usually requires thousands of servers. Testing new chips or system layouts in the real world takes a lot of time and money. The new simulator solves this by copying tasks like data processing, request sorting, and memory use on a regular computer.
The technology goes beyond traditional graphics processing units to support newer chips. Users can test neural processing units and memory chips that handle calculations directly within a virtual data center. This helps researchers see if a certain chip will speed up processing, use less power, and stay stable across huge server networks. It can also analyze setups where different server resources are physically separated but connected over a network.
The research was led by master's students Cho Jae-hong and Choi Hyun-min. It won the best paper award at the IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) 2026. The simulation tool has been released to the public as open-source software so businesses and schools can use it freely.
"AI service competitiveness is determined not only by the model itself but also by the infrastructure technology that operates it stably and efficiently," Park said. "I hope this simulator will serve as an important foundation for researchers and the industry to develop next-generation AI infrastructure more quickly and efficiently."
(Reference Information)
Journal/Source: IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) 2026
Title: LLMServingSim 2.0: A Unified Simulator for Heterogeneous and Disaggregated LLM Serving Infrastructure
Link/DOI: 10.1109/ISPASS69572.2026.00012
Copyright ⓒ Aju Press All rights reserved.