CFD on GPUs in Aerospace Applications. Benchmarking the Fluent native GPU solver on aerospace applications, and how to approach purchasing GPUs for CFD as a business case.

Abstract

Running CFD simulations on GPUs is becoming commercially viable, and one major early adopter is the automotive industry, where external aerodynamics cases can be run at a fraction of the time compared to simulations on CPUs. The aerospace industry has not yet adopted GPUs to the same extent, as aerospace cases often require support for more complex physics such as compressible flows and combustion. This study compares the performance of the Ansys Fluent GPU solver with the CPU solver in aerospace applications and is carried out with the support of GKN Aerospace and EDR & Medeso, a reseller of Ansys software. It uses a novel approach to evaluate the attractiveness of purchasing GPUs for a local cluster, compared to purchasing CPUs, from a cost, power consumption, and strategic perspective. A case-based methodology is used to compare the solvers with 3 simulation setups that are representative of typical aerospace applications. The current version of the GPU solver supports all the necessary features to run 2/3 cases, although it requires minor simplifications to the case setups. For the cases it does support, key results include GPU simulations providing a time reduction of 41-98% per iteration, an energy consumption reduction of 88-93% per iteration, a 27-73% reduction in iterations to reach convergence, a cloud computing cost reduction of 83-91% and a total cost of ownership reduction of 48-67% for systems with equivalent simulation capacity on a local cluster. If the simulation capacity demand for simulation setups that the GPU solver supports is sufficient, purchasing GPUs for CFD simulation is a cost-effective and energy-efficient solution to meet simulation capacity demands in comparison to purchasing CPUs. The speedups provided by the Ansys Fluent GPU solver can be leveraged to generate significant value in an engineering process by enabling more design iterations, improved simulation fidelity, and faster simulation turnaround, compared to the CPU solver.