The mechanical properties of metal additive manufacturing products are strongly influenced by the morphology of the solidification microstructure formed during rapid solidification in the molten pool created by beam irradiation. Within the molten pool, steep temperature gradients and high cooling rates lead to rapid solidification rates reaching up to ~1 m/s. To accurately predict the microstructure formed under such far-from-equilibrium conditions, it is essential to account for non-equilibrium effects like solute trapping at the solid-liquid interface . The phase-field (PF) method has been developed as a powerful approach for simulating solidification microstructure, including dendrite growth, and by employing a nanometer-scale PF interface width, it can quantitatively reproduce these non-equilibrium phenomena . Ji et al. extended this approach and successfully reproduced the formation of banded microstructures—characteristic of rapid solidification, where planar interface growth alternates with dendritic growth—through PF simulations . However, the nanometer-scale grid resolution results in extremely high computational costs, limiting three-dimensional (3D) simulations to only a single dendrite arm domains . In this study, we aim to accelerate PF computation to enable practical 3D simulations of rapid solidification. First, nonlinear preconditioning is applied to the PF variables to enable stable simulations even on coarse grids. In addition, we implement a block-structured adaptive mesh refinement (AMR) method, which dynamically allocates fine grids around the PF interface and coarse grids elsewhere thereby improving computational efficiency. These simulations are executed in parallel on multiple GPUs to further accelerate computations. To demonstrate the effectiveness of the developed method for 3D simulations, we reproduce banded structures in a 3D system and evaluate the influence of temperature conditions on their microstructural morphology.