Abstract:The real-time, continuous and rapid arrival properties of data streams decide the real-time processing capability of data stream. Quantiles are commonly used for describing data stream with low dimension distribution. The research focused on mining powerful computing capacity and high memory bandwidth of Graphics Processing Unit (GPU) to compute data stream quantiles, and presented a GPU cooperated parallel processing model of data stream based on Computing Unified Device Architecture (CUDA) as well as parallel computing method of data stream quantiles which increased data stream processing bandwidth remarkably with precision no less than pure CPU algorithm.