Developing synergy between styling design and aerodynamic efficiency has been an ongoing challenge in the automotive industry for many years. Designers and aerodynamicists have to meet strict deadlines and have limited resources to iterate and evolve the design and performance of the vehicle exterior during early concept development. The number of possible styling variants and configurations can often reach into the hundreds. Physical wind tunnel measurements are simply not practical from a time or cost perspective to evaluate these designs. Therefore, increased performance and accuracy of Computational Fluid Dynamics (CFD) simulations have become the main objective of every vehicle OEMs. Today, the GPU hardware, particularly the memory and performance, has reached a point where there is an increased interest in their usage for aerodynamics simulations. GPUs offer potential simulation performance improvements due to lower power consumption and hardware costs. With this in mind, a GPU based version of the Lattice Boltzmann Method (LBM) fluid solver has been recently implemented. LBM is known to be well suited for GPU computation as the method is naturally vectorizable. Additionally, as a low dissipation, inherently transient approach, it is also well-suited for accurate evaluation of external automotive aerodynamics.
This study provides a comparison between the relative performances of GPU based LBM vs its CPU based equivalent using the publicly available DrivAer aerodynamic model. A comparison of turn-around time on NVIDIA V100 and A100 cards vs an Intel Skylake CPU cluster is demonstrated. The resulting flow structures and forces will then be compared to show that the results are numerically equivalent, with significant performance gains on the GPU. Different models with different balances between surface elements and fluid nodes will also be evaluated, allowing for the relative computational cost of these different simulation elements to be discussed.