| 논문명 |
AI 데이터센터 직접액체냉각 적용을 위한 설비인프라 모듈계획 및 에너지효율 정량화 연구 / Design Strategies for Cooling Infrastructure Modules and Energy Performance Assessment of Direct Liquid Cooling in AI Data Centers |
| 수록사항 |
대한건축학회논문집, Vol.42 No.7 (2026-07) |
| 주제어 |
AI 데이터센터; 직접액체냉각; 냉각 모듈 및 설비인프라; 에너지효율 평가; 전력사용효율 ; AI Data Center; Direct Liquid Cooling; Cooling Module & Infrastructure; Energy Efficiency; Power Usage Effectiveness (PUE) |
| 요약1 |
본 연구에서는 고밀도 AI 서버 환경을 대상으로 액체냉각 기술 적용이 가능한 데이터센터 냉각 인프라 모듈 설계 방안을 제시하고, 냉각 방식에 따른 공간 구성과 에너지 성능을 정량적으로 비교하였다. 이를 위해 10 MW급 도심형 엣지 데이터센터를 기준으로 기존 공랭식 FWU 시스템, 리트로핏 적용이 가능한 1상 직접액체냉각(1P DLC) 시스템, 그리고 신규 구축을 위한 2상 직접액체냉각(P2P DLC) 시스템의 세 가지 냉각 시나리오를 설정하고 냉각 인프라 구조와 에너지 성능을 분석하였다. |
| 요약2 |
This study investigates the design strategy and energy performance of direct liquid cooling (DLC) infrastructure for high-density AI data centers. As AI accelerators and high-performance computing systems increase rack-level power density, conventional air-cooling systems face limitations in cooling capacity and energy efficiency. To address these challenges, this study proposes a modular cooling infrastructure architecture based on chip-level cold plate cooling and evaluates its performance under different cooling configurations. Three cooling scenarios were considered for a 10 MW edge data center: a conventional air-cooling system using fan wall units (FWU), a retrofit liquid cooling system combining single-phase DLC with rear-door heat exchangers (1P DLC + RDHx), and a two-phase direct-to-chip system (P2P DLC). The cooling infrastructure configuration and energy consumption of each system were analyzed using system-level energy modeling. The results show that the required white space decreases from approximately 3,000 m2 for the air-cooling system to 300 m2 for the 1P DLC system and 120 m2 for the P2P DLC system. In addition, the PUE improved from 1.305 to 1.104 and 1.082, respectively. These results indicate that liquid-based cooling systems can significantly improve both spatial efficiency and energy performance in high-density AI data centers. |