Experiences gained when porting a production finite difference code written in Fortran to GPUs using OpenACC are summarized. While the original code was written taken into consideration vectorization and shared-memory parallelization via OpenMP, the loop structures and the proper placement of directives had to be analyzed in depth, leading to a partial recoding of core loops that in some cases led to surprising results. As in all applications running on GPUs, minimization of transfer between the CPU and GPU was of paramount importance. Simulations run to date indicate that 1 Nvidia VT100 is equivalent to approximately 50 Xeon CPU cores for simple geometries and more than 30 Xeon CPU cores for complex cases. Besides some classic testcases, a complex industrial case (external aerodynamics for a generic Volkswagen Golf VII) is also included.