Abstract: As scales of radio interferometer telescope arrays increase, it becomes a tremendous challenge for the traditional solution to deal with intense real-time radio-astronomy data with high performances and low costs. In this paper we propose a creative and effective GPU-based solution to fit the needs of processing data from new arrays. We first give a brief description of the requirements on an effective solution, i.e. a high-speed low-cost procedure to transmit data and compute for real-time processing of data from a large-scale radio interferometer. We then present our GPU-based model, which improves practical computing performances of a single GPU card through accurately mapping data blocks in cross-correlation computations to GPU grid threads. The model achieves performances of up to 77% of those theoretically possible. Finally, several tests are carried by tailoring the model parameters to the requirements of the Tianlai project. The tests pave the way for realizing an on-line signal processing system with even much higher performances and a lower cost in an environment of GPU clusters. Such a system is to work for future more complicated and much larger radio interferometers.