Computer architecture simulators are widely used to explore new ideas. The accuracy of the simulator affects the credibility of the research results. However, there is a lack of studies to evaluate the error of simulators with respect to recent processors. In this paper, we use a recent ARM server as the target machine and configure the gem5 simulator as close as possible with official public information and measurement data. In single-core mode, we perform accuracy evaluation using SPEC CPU 2006, SPEC CPU 2017, PARSEC, and SPLASH-2x benchmarks with a mean absolute percentage error (MAPE) of 26.31%. In multi-core mode, the MAPE stays around 30% for PARSEC and SPLASH-2x as the core count varies. These results help architecture researchers to quantitatively understand the performance errors between the gem5 simulator and the modern ARM processor.