Abstract: The thrashing of pixel cache in GPU is frequent, which greatly decrease the performance of GPU. Based on the research of the parallelization characteristics of multi data stream processing in GPU, a pixel write merging buffer technique is proposed. The cache access is merged according to the address before. It can reduce the number of cache access, reduce the cache thrashing, and improve the performance of the cache. The pixel write combining buffer technology can reduce the average cache trash by 60% and the maximum reduction of 70%. The result shows that the write merge buffer technology of pixel cache can reduce the cache thrash, and is very suitable for the design of the pixel cache of the embedded graphic processor.