The article is devoted to the modern direction of information technologies – parallel computing platform CUDA invented by NVIDIA. The work describes the possibility of data-flow computing on graphics processors. A model of parallel programming based on low-level platform named CUDA is presented. The differences in the organization of data processing threads and memory access using CPU or GPU are highlighted. Effective GPU programming requires thorough understanding of the principles of parallel programming as well as models of overlapping, data exchange and knowledge of various architectural restrictions of these processors. It is showed that the effectiveness of CUDA technology on performance increasing is relies on the possibility of parallelizing code. Multithreaded exchange of information between the CPU and GPU is inspected and its effectiveness in various aspects of data processing is analyzed. The novelty technological features of state-of-the-art processor named NVIDIA Kepler are listed.