Considerations for GPU Acceleration of Parallel CFD

Increased levels of parallel CFD by utilizing graphical processing units (GPUs) in a high performance computing (HPC) environment can enabled larger and complex CFD simulations to be addressed in a practical and cost-effective way as needs grow for transients, high- resolution LES, and multidiscipline simulations. This paper examines the current state of parallel CFD for industry use, and opportunities for a second level of CFD solver parallelism through hybrid CPU-GPU heterogeneous computing. GPUs are developed and deployed to share computational tasks with the CPU, in particular computational tasks that benefit from massively-parallel shared memory numerical operations. The heterogeneous benefits to high fidelity CFD simulations and their HPC efficiency will continue to grow requirements for balanced co-processing involving large GPU-based servers as a foundation for future distributed memory clusters.