Smoothed Particle Hydrodynamics (SPH) is a fully Lagrangian, particle-based technique for fluid-flow modeling. As a gridless method, it appears to be a natural approach to simulate multi-phase flow with complex geometries. Since SPH involves a large set of short-range particle-particle interactions, numerical implementations present a high degree of spatial data locality and a significant number of independent computations. Therefore, the numerical code can be easily written in a massively parallel manner. The main purpose of this study is to discuss the issues related to the implementation of the SPH method for computation using Graphics Processing Units (GPU). The study is supported by two-dimensional validation cases: the lid-driven cavity and oscillation of a droplet. The obtained results show a good accuracy of the method, as well as, high numerical efficiency of its GPU implementation.