Acceleration of the finite element method using hybrid OpenMP-CUDA

Graphics processing units (GPUs) are efficient in accelerating algorithms that are highly parallelizable. This paper discusses various aspects of parallelizing the traditional finite element algorithm, whose communication-intensive nature makes it difficult to be parallelized in a straightforward manner, and proposes solutions to alleviate the acceleration bottlenecks. The examples show that decent speedup can still be achieved over OpenMP-enabled CPUs.