GPU implementation of lattice Boltzmann method for flows with curved boundaries

An efficient implementation of the lattice Boltzmann method (LBM) using the compute unified device architecture (CUDA) provided by nVidia was presented to simulate flows with curved boundaries. The flow around a circular cylinder was investigated as a typical case and satisfactory results in terms of precision and performance were obtained. The predicted drag coefficient (CDCD), lift coefficient (ClCl) and Strouhal number (StSt) agreed well with the results of previous related studies. The test results indicate that the number of threads has a great influence on the performance of graphical processing unit (GPU) based implementation. A total of 128 threads per block can lead to the best performance with nearly an 18-fold speed increase.