In this paper a parallelization of a Shallow Water numerical scheme suitable for Graphics Processor Unit (GPU) architectures under the NVIDIA™'s Compute Unified Device Architecture (CUDA) framework is presented. In order to provide robust and accurate simulations of real flood events, the system features a state-of-the-art Finite Volume explicit discretization technique which is well balanced, second order accurate and based on positive depth reconstruction. The model is based on a Cartesian grid and boundary conditions are implemented by means of the implicit local ghost cell approach, which enables the discretization of a broad spectrum of boundary conditions including inflow/outflow conditions. A novel and efficient Block Deactivation Optimization procedure has also been adopted, in order to increase the efficiency of the numerical scheme in the presence of wetting-drying fronts. This led to speedups of two orders of magnitude with respect to a single-core CPU. The code has been validated against several severe benchmark test cases, and its capability of producing accurate fast simulations (with high ratios between physical and computing times) for different real world cases has been shown. •Shallow Water Equations parallel numerical scheme suitable for GPU.•Speedup of two order of magnitude over a single-core CPU.•Robust treatment of wetting and drying fronts.•Optimization strategies for partially dry domains.•Real-life simulations with open and closed boundaries.