Numerical study on the effect of liquid flow on three-dimensional dendrite growth is still a challenging topic. Herein, high-performance phase–field lattice Boltzmann (PF-LB) simulations were performed to investigate the effect of natural convection on dendrite morphology and the possibility for causing fragmentation. Parallel computing in multiple graphics processing units (GPUs) with dynamic load balancing for the block-structured adaptive mesh refinement (AMR) scheme (parallel-GPU AMR) was applied to the PF-LB simulations as a high-performance computing tool in a GPU supercomputer. Parallel-GPU AMR PF-LB simulations showed that the growth of dendrites with natural convection in two and three dimensions were quite different. The dendrite tip velocity increased in the following order: upward buoyancy, no gravity, and downward buoyancy. Downward and upward buoyancy enhanced and restricted the growth of the secondary arms, respectively. The root size of the secondary arms growing from the bottom was drastically affected by the flow direction. However, the dendrite fragmentations were not observed in the present simulations.