Recent multi-core paradigm shift in processor architecture including CPU, GPU and APU have provided the industry with new tremendous performance improvement possibilities in harnessing their parallelization capabilities. At the same time, Smart Grid concept introduced new challenges for Distribution Management Systems with demands for modeling low voltage level with distributed generation, increasing the amount of data by several fold. This paper presents a research into the parallelization of the real-time three-phase distribution power flow calculation on the Graphic Processing Units (GPUs). The Parallelization is implemented based on the Forward Backward algorithm by parallelizing radial lateral calculation of voltages in forward sweep and branch currents in backward sweep over parallelization layers. Special GPU kernel is developed using the OpenCL standard and language. High dependence on GPU kernel memory management is analyzed for the given implementation. The results are provided showing the comparison between CPU and GPU implementation.