GPU Accelerated Iterative Solutions Of Large-Scale Geotechnical Finite Element Computations

In geotechnical finite element computations, a series of resulting linear systems of equations are required to solve. With the increasing of problem size and computational complexity, diversification of calculation parameters, as well as the computational accuracy, the computational work, the required computer storage and the computer runtime may increase rapidly. Serial computing platform and available computational methods may not satisfy the computing demand, so that the parallel strategy has to be adopted. The graphics processor unit, firstly employed for graphical processing, has a good computing performance in high-intensity floating point operations, memory bandwidth and parallel implementation, especially the programmability. Hence, general-purpose GPU is becoming a hot field with various applications.The purpose of this paper is taking GPU parallel computing platforms into computing geotechnical finite element problem, and accelerating SpMV computing by GPU card.3D Biot’s consolidation is written on PGI Fortran compiler using Fortran language, the compute result is checked by Abasqus to verify the correctness; two different geotechnical problems are analysed.(1)3D Biot’s consolidation is a typical geotechnical problem, Biot’s equations can become systems of linear equations though the finite element discretization, Generalized Jacobi(GJ) preconditioned SQMR iterative method is applied based on CPU and CPU-GPU platform. On the computing scale side, sparse matrix storage method is used and the maximum freedom degrees can be as much as2.8million on PC; On the computing efficiency side, maximum double precision computing speedup can reach2.79based on GPU parallel acceleration.(2) There is a significant material stiffness contrast for soil-structure interaction problems. Though piled-raft foundation consolidation and drained problems, preconditioning techniques which were recently proposed for soil and structure interaction problems are compared based on CPU and CPU-GPU platforms, preconditioners computational performance are investigated and GPU accelerating effects are evaluated.