Large scale parallelisation of the material point method with multiple GPUs

The material point method (MPM), which is a combination of the finite element and meshfree methods, suffers from significant computational workload due to the fine mesh required in spite of its advantages in simulating large deformations. This paper presents a parallel computing strategy for the MPM with multiple Graphics Processing Units (GPUs) to boost the method's computational efficiency in large scale problems. Domain decomposition method is used to split the workload over subdomains onto a number of GPUs. Communication between the subdomains is implemented by the data transfer between GPUs and random access memory. On each GPU the MPM algorithm is parallelised over nodes or particles. Benchmark problems of slurry runout and surface pipe penetration are analysed to quantify the speedup of the multiple-GPU parallel simulations over the sequential counterparts on the central processing unit. The maximum speedup with 1 GPU is 84 and increases to ∼1280 using 16 GPUs.