high performance computing on graphics processing units: hgpu.org

Newmark family of algorithms have been utilized by many engineering applications for the solution of nonlinear dynamic analysis of various structural models. Dynamic and nonlinear nature of such problems and numerical stability requirements of the algorithms increase the need for computation power in order to achieve practical solution times. Thus, this study intends to decrease the analysis time for nonlinear dynamic analysis of large scale structural models utilizing the GPUs. In the implementation, explicit version of the Newmark family of algorithms was utilized. This type of algorithm enabled the computations to be applied on each finite element eliminating the need for global matrix assembly. Two different GPU implementations were tested. In the first approach, creation of elemental matrices and computation of the explicit Newmark algorithm are separated into two different kernels. The second approach fuses these two kernels at compile time into a single kernel code. Both implementations were developed using CUDA language. Implementation details of both algorithms were discussed in detail noticing optimization differences. Both GPU implementations were tested and compared with a CPU implementation using models with varying sizes.