Finite particle method (FPM), as a novel numerical method, has advantages in dealing with complex behaviors of structures due to its unique concepts of point value description, path unit description and fictitious motion. The prominent advantage of this method is that nor the integration and inversion of stiffness matrix or the nonlinear iteration is involved during the whole calculation process, and also the governing equations of particles and element forces are solved independently as well, that means the whole calculation process is decoupled, and the solution complexity just increases linearly with the model scale. These make it possible to significantly improve the solution efficiency by the high-performance parallel computing technique. Combining the theory of the FPM and high-performance computing technology, this study explores the feasibility of using the finite particle method of GPU parallel framework to carry out fast calculation of large-scale structure models. According to the unified framework and computational flow of the FPM, the computational data storage and thread mapping mechanism is established firstly, and then the parallel acceleration of the whole calculation process including particle force calculation, integration and solution of governing equation is realized. Furthermore, this study mainly focuses on the calculation efficiency of the self-developed C++ program employing GPU parallel computing framework. Numerical results for testing the acceleration performance show that the calculation efficiency of beam, shell and solid can be improved by dozens of times for ordinary desktop computers. Moreover, it has certain advantages compared with the CPU parallel architecture in ABAQUS using different numbers of processing cores. By combining the GPU parallel acceleration technology and MPI distributed technology, the implementation scheme of particle partition and data synchronization between them is formulated, and the parallel acceleration strategy of GPU cluster is established to further improve the parallelism and computable scale, optimize the computational efficiency, and provide the capability to perform high-efficiency numerical simulations for city-block-scale buildings.