This paper addresses an approach that uses GPU (graphic processing unit)-based processing architecture model and its parallel algorithm for high-dimensional data streams over the irregular streams in order to satisfy the real-time requirement under the resource-constraints. This six layers model combines the GPU high wide-band property of data processing with analysis data stream in a sliding window. Next, canonical correlation analysis is carried out between two high-dimensional data streams, by a data cube pattern, and a dimensionality-reduction method in this framework based on compute unified device architecture (CUDA). The theoretical analysis and experimental results show that the parallel processing method can detect correlations on high dimension data streams, online, accurately in the synchronous sliding window mode. According to the pure CPU method, this technique has significant speed advantage and conducts the real-time requirement of highdimensional data stream very well. It provides a common strategy for the applied field of data stream mining.