This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Due to the significant benefits being realized in time and cost to solution, the use of General Purpose Graphical Processing Units (GPGPUs) over traditional Central Processing Units (CPUs) is becoming ever greater in the industrial Computational Fluid Dynamics (CFD) world. As the processing architecture of a GPGPU is fundamentally different to a CPU, alongside arises a need to optimize and accelerate solver performance specifically on GPGPUs. While certain processes become less expensive as a result of GPU acceleration, certain other processes which were relatively quick on the CPU, may become dominant portions of the overall run time on GPU. For unsteady CFD problems involving rigid body motion, the costs of boundary interface computations and the maintenance of a dynamic framework for sliding meshes can oftentimes become one such performance bottleneck on GPGPUs. In this paper, Boundary Interface Caching (BIC) is presented as a method to accelerate overall solver performance by significantly cutting down on these over-head processing costs for sliding mesh / moving mesh CFD simulations. The fundamental methodology of boundary interface caching is firstly explained, following which the Simcenter STAR-CCM+ solver is used to demonstrate the benefits of boundary caching on two industrial use cases: (1) an external aerodynamics simulation of a sports car, and (2) an acoustic simulation of a HVAC (Heating, Ventilation, Air-conditioning) fan. Both simulations are first run without boundary interface caching to establish a baseline. They are subsequently run with boundary interface caching, and this is done on both CPU and GPU architecture. The resulting convergence behavior, accuracy of solution, and solver performance are presented, compared and contrasted in this paper. For STAR-CCM+ simulations with rigid body motion, it is demonstrated that our proprietary boundary caching algorithm provides significant improvements in overall solver time on GPGPUs, whilst maintaining similar levels of accuracy as the traditional CPU based simulations.