In order to make the numerical model adapt to the fast development trend of high performance computing with hybrid architecture, based on the OpenACC, this paper presents the GPU accelerated optimization test for the three-segment procedure which is Beijing Climate Center Atmospheric General Circulation Model version 3.0 (BCC_AGCM3.0) in dynamic part. By means of asynchronous execution settings, cyclic inward migration, data management and vector parameterized configuration, we parallel the GPU acceleration of the computationally intensive part of the program segments in model, compare the optimization efficiency and validate the correctness. The experimental results show that the efficiency of the three-stage GPU acceleration in BCC_AGCM climate model, is increased by more than three times and the relative error of global vorticity root mean square (RMS) is controlled within a certain range. The acceleration methods and strategies are valuable for transplanting and optimizing numerical weather and climate models in heterogeneous environments.