This paper presents the development of algorithms and their implementation in a graphics processing unit (GPU)-accelerated computational fluid dynamics (CFD) solver for efficient steady and unsteady analyses of flowfields within multirow turbomachinery. To enhance convergence, the data-parallel lower–upper relaxation method is incorporated as a residual smoother within the Runge–Kutta framework. The method allows for a large time step while maintaining parallelism at a cell level. The mixing plane and sliding plane methods are implemented for multirow turbomachinery scenarios. Particular attention is paid to the implementation and optimization of these two rotor–stator coupling methods within the Compute Unified Device Architecture. Various strategies are employed to leverage the capabilities of GPUs for high-performance computation. The presented solver achieves speedup factors of about 18 and 24 for single- and double-precision floating-point arithmetic, respectively, in GPU computing under equivalent thermal power consumption to that in central processing unit computing. Parallel scalability tests on an eight-GPU cluster demonstrate high parallel efficiency, reaching 96% and 98% for strong and weak scalings, respectively. For validation, a series of steady analyses are conducted to analyze the flowfields within a fan stage and an axial compressor, demonstrating good agreement of the overall performance metrics and radial flow profiles with the corresponding experimental data. Furthermore, the stall behavior of the compressor at 65% design speed is analyzed, and the results are consistent with the experimental findings.
Wang et al. (Thu,) studied this question.