LAUSR

Efficient computation of density-based topology optimization using GPU computing is proposed.Workload of CUDA threads is properly balanced by the multi-granular GPU implementation of computationally intensive tasks.The use of on-chip memory is maximized in matrix-vector product operations using 3D CUDA kernels.The speedup and wall-clock time of the solving stage are evaluated using Jacobi and geometric multigrid preconditioning techniques.Numerical experiments show that the GPU instance of PCG using geometric multigrid preconditioner provides good performance results. This paper proposes a well-suited strategy for High Performance Computing (HPC) of density-based topology optimization using Graphics Processing Units (GPUs). Such a strategy takes advantage of Massively Parallel Processing (MPP) architectures to overcome the computationally demanding procedures of density-based topology design, both in terms of memory consumption and processing time. This is done exploiting data locality and minimizing both memory consumption and data transfers. The proposed GPU instance makes use of different granularities for the topology optimization pipeline, which are selected to properly balance the workload between the threads exploiting the parallelization potential of massively parallel architectures. The performance of the fine-grained GPU instance of the solving stage is evaluated using two preconditioning techniques. The proposal is also compared with the classical CPU implementation for diverse topology optimization problems, including stiffness maximization, heat sink design and compliant mechanism design.