Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Numerical Verification of a Parallelized Code Run on GPU Applied to Moving Heat Source Autogenous Welding Processes
Submission Author:
Ernandes José Gonçalves do Nascimento , SP
Co-Authors:
Ernandes José Gonçalves do Nascimento, Arthur Mendonça de Azevedo, Elisan dos Santos Magalhães, Luiz Eduardo dos Santos Paes
Presenter: Ernandes José Gonçalves do Nascimento
doi://10.26678/ABCM.COBEM2023.COB2023-0222
Abstract
The industrial revolution fostered by the application of the Computational Fluid Dynamics (CFD) techniques had made possible a series of enhancements in modern engineering design. The recent advancements in processing hardware gave room for the modeling and simulation of a variety of higher complexity physical phenomena. However, although there has been a recent significative increase in hardware parallel computing performance, most CFD codes still make use of a very limited serial-based processing methodology. The arithmetic operations accuracy and the execution advantages of developing a numerical solution through parallel Graphics Processing Unit (GPU) computing has been put in question for years in the scientific community. To fulfill this gap, a numerical verification of an inhouse parallelized (Compute Unified Device Architecture) CUDA-C language code was performed. A moving heat source autogenous welding process was simulated to compare the GPU processed data to a commercial code solution. The two analyses were performed by applying the Finite Volume Method (FVM) to solve the transient heat conduction Partial Differential Equation (PDE) in a fixed structured uniform mesh. A first order temporal discretization scheme and temperature dependent non-linear thermal properties were applied in both solutions. The melting and solidification processes were accounted for by the application of the enthalpy method. The moving heat source was modeled through a time and space dependent Gaussian whole conical volumetric heat distribution profile. A Nvidia GeForce RTX 3090 graphics card with 24 GB of dedicated memory was applied as the inhouse code processing hardware. The commercial code solution was run on an Intel® Core™ I7 11700K with a 3.6 GHz base clock and 16 parallel processing threads. The research outcomings suggested that GPU processing deliveries a similar computational accuracy to that of a CPU but a much higher efficiency in hardware usage. The results also evidenced that CPU parallelization does have an optimum number of simultaneously run processes that is far below its hardware maximum number of threads.
Keywords
numerical heat transfer, Finite volume method, CUDA-C language, LASER welding process, Numerical phase change
