Efficient isogeometric topology optimization via multi-GPUs and CPUs heterogeneous architecture

Isogeometric analysis (IGA) has higher continuity and data density than finite element method (FEM) at the same order, but the inefficient serial computation method will limit the model size of IGA. In this paper, a novel scalable multi-GPU and multi-core CPU heterogeneous parallel scheme for large-scale IGA-based topology optimization (ITO) is proposed, encompassing a multi-GPUs/CPUs parallel solving algorithm, collaborative stiffness assemble and sensitivity analysis parallel scheme, and performance optimization scheme at hardware-level. Two efficient algebraic multigrid (AMG) methods are compared with the proposed parallel solution strategy, and the performance of parallel ITO scheme in this paper is evaluated with different computing configurations. GPU parallel scheme are two orders of magnitude faster compared to parallel CPU, allowing a single computing node to solve an ITO problem with 5.78 million DOFs (Degree of Freedoms) in 46 s.