DescriptionThe conjugate gradient method with multigrid preconditioners (MGCG) is used in scientific applications because of its high performance and scalability with many computational nodes. GPUs are thought to be good candidates for accelerating such applications with many meshes where an MGCG solver could show high performance. No previous studies have evaluated and discussed the numerical character of an MGCG solver on GPUs. Consequently, we have implemented and optimized our “kinaco” numerical ocean model with an MGCG solver on GPUs. We evaluated its performance and discussed inter-GPU communications on a coarse grid on which GPUs could be intrinsically problematic. We achieved 3.9 times speedup compared to CPUs and learned how inter-GPU communications depended on the number of GPUs and the aggregation level of information in a multigrid method.