The interface thickness of phase-field model limits its simulation scale. In order to shorten the computation time, the large undercooling degree higher than that under the conventional casting process condition is often employed in the three-dimensional dendritic growth simulation, and many qualitative simulations are mainly aimed at a single dendrite and multi-dendrites morphologies for pure substances or binary alloys. The problems of low computational efficiency, small-scale simulation and limited to qualitative research exist on a single CPU computation using phase-field method. To simulate microstructure evolution process during actual casting solidification more effectively, a high performance computing method based on CUDA+GPU architecture is explored in this paper, and larger-scale computation is implemented by the concurrent execution of multiple threads to improve computational efficiency. The three-dimensional dendritic growth of pure SCN is quantitatively simulated on a single GPU, and the computational efficiency based on a single GPU and a single CPU is also compared under the same condition. The simulation results show that a speedup of 98.52 is achieved when the grid size is 5123 on the GPU, with computational efficiency being greatly improved. Meanwhile, the calculated values of the dendritic tip velocities and the tip radius on the GPU are identical with the values of the microscopic solvability theory and the references’ simulation results, which validates the GPU parallel algorithm.