Multi-GPU Accelerated Large-scale Phase-field Simulations for Dendritic Solidification
Tomohiro Takaki, Kyoto Institute of Technology
The phase-field method is widely accepted as the most powerful and accurate numerical model for expressing dendrite growth during alloy solidification. However, since the computational cost is high due to the diffuse interface model, dendrite growth simulations by the phase-field method have been conducted only for limited space conditions, such as two-dimension or single dendrite in three-dimension. Since solidification structures are formed through competitive growth of multiple dendrites, it is essential for accurate prediction of solidification microstructures to perform large-scale phase-field simulations efficiently. In addition, in alloy solidification under terrestrial gravity, liquid flow necessarily occurs. Therefore, coupling simulations of phase-field and liquid flow are also important for the accurate prediction of solidification structures. However, the computational cost increases more. With these points in consideration, we have been performing large-scale phase-field simulations of multiple dendrite growth and dendrite growth with liquid flow by parallelizing multiple graphical processing units (GPU). General-purpose computing on GPU has emerged as a high-speed computing scheme. Meanwhile, since the memory size of the GPU is small compared to that of a CPU, we need parallel computing with multiple GPUs for the large-scale simulation. Therefore, we have been using a GPU-rich supercomputer, TSUBAME, at Tokyo Institute of Technology. Thus far, we have enabled the large-scale simulations of the competitive growth of multiple dendrites during directional solidification of a binary alloy [Acta Mater. 118 (2016) 230-243, ISIJ Int. 56 (2016) 1427-1435, Materialia 1 (2018) 104-113], three-dimensional dendrite growth in liquid flow [J. Crystal Growth 4474 (2017) 154-159, J. Crystal Growth 483 (2018) 147-155], and the growth of multiple dendrites with motion [Comp. Mater. Sci. 1447 (2018) 124-131]. In this talk, our studies regarding dendrite growth using large-scale phase-filed simulations on the GPU supercomputer are introduced.
Tomohiro Takaki, Kyoto Institute of Technology
The phase-field method is widely accepted as the most powerful and accurate numerical model for expressing dendrite growth during alloy solidification. However, since the computational cost is high due to the diffuse interface model, dendrite growth simulations by the phase-field method have been conducted only for limited space conditions, such as two-dimension or single dendrite in three-dimension. Since solidification structures are formed through competitive growth of multiple dendrites, it is essential for accurate prediction of solidification microstructures to perform large-scale phase-field simulations efficiently. In addition, in alloy solidification under terrestrial gravity, liquid flow necessarily occurs. Therefore, coupling simulations of phase-field and liquid flow are also important for the accurate prediction of solidification structures. However, the computational cost increases more. With these points in consideration, we have been performing large-scale phase-field simulations of multiple dendrite growth and dendrite growth with liquid flow by parallelizing multiple graphical processing units (GPU). General-purpose computing on GPU has emerged as a high-speed computing scheme. Meanwhile, since the memory size of the GPU is small compared to that of a CPU, we need parallel computing with multiple GPUs for the large-scale simulation. Therefore, we have been using a GPU-rich supercomputer, TSUBAME, at Tokyo Institute of Technology. Thus far, we have enabled the large-scale simulations of the competitive growth of multiple dendrites during directional solidification of a binary alloy [Acta Mater. 118 (2016) 230-243, ISIJ Int. 56 (2016) 1427-1435, Materialia 1 (2018) 104-113], three-dimensional dendrite growth in liquid flow [J. Crystal Growth 4474 (2017) 154-159, J. Crystal Growth 483 (2018) 147-155], and the growth of multiple dendrites with motion [Comp. Mater. Sci. 1447 (2018) 124-131]. In this talk, our studies regarding dendrite growth using large-scale phase-filed simulations on the GPU supercomputer are introduced.
