Interest in scientific programming using graphics processing units (GPUs) has exploded in recent years. The advent of NVIDIA's CUDA programming language in early 2007 enabled GPU acceleration of numerical software to become mainstream. Relative to central processing units (CPUs), these devices have extremely high floating point operation capability and memory bandwidth. Combined with relatively low cost, they are attractive alternatives to more expensive traditional supercomputers.

Porting existing computational fluid dynamics methods to the new hardware is not always straightforward. Modern GPUs are massively parallel, some consisting of over 400 processors, utilizing a unique hierarchy of computational units and memory management. Fully exploiting this architecture for CFD solvers requires the development of new algorithms tailored to the devices. To that end, this work presents a solution method for the Navier-Stokes equations using the SIMPLER algorithm on structured Cartesian grids. A block-iterative scheme with a parallel recursive tridiagonal solver is used for the discretized equations, giving considerable performance advantages over prior point-iterative implementations. Using a $200 GPU in a standard workstation, accelerations of over 20x are observed compared to a serial CPU implementation for rotorcraft simulations.