Numerical Investigation of the Thermal Environment in a Data Center Under Dynamic Server Loads Driven by Artificial Neural Network Models | CU Experts

This study numerically investigates the transient thermal response of a raised-floor data center under dynamic AI server loads. Power profiles of DGX A100 servers running GEMM, VGG-19, and ResNet-152 workloads were used as time-varying boundary conditions in CFD simulations. Six load scenarios were analyzed to evaluate the dynamic responses of cooling capacity, rack inlet and outlet temperatures, and net airflow rate. The results show that the response lag of air-conditioning units causes a temporary mismatch between cooling supply and server heat dissipation, leading to transient inlet-temperature fluctuations. Fluctuating workloads produce larger outlet-temperature amplitudes and higher temperature change rates than stable workloads, while mixed workloads change the overall fluctuation intensity and affect neighboring-rack airflow through cold-aisle pressure interactions. These findings can support workload-aware rack arrangement and adaptive cooling control for future AI-oriented data centers.

VIVO