Numerical simulation of turbulent flows using the Finite Element Method and GPU-CUDA parallelization

Resumo

In the present work, an investigation is carried out on the use of Graphics Processing Units (GPUs) as
a powerful tool to accelerate Computational Fluid Dynamics (CFD) simulations and problems of interest for CWE
(Computational Wind Engineering). With this intent, different code versions (OpenMP and CUDA-Fortran) are

utilized to compare response accuracy and computational performance obtained using an explicit two-step Taylor-
Galerkin scheme, with turbulence modeling via Large Eddy Simulation (LES), where the flow equations are

discretized in the context of the Finite Element Method (FEM) considering eight-node hexahedral elements. From
the 3D simulation investigated here, it is observed that the use of GPUs leads to high levels of computational
processing speed, generating little, if any, loss of accuracy in the numerical results, and still allows increasing the
refinement level of computational meshes and time increment without increasing the simulation time compared to
other approaches.