Abstract: We present an efficient model to analyze and improve the performance of general-purpose computation on graphics processing units (GPGPU). After analyzing the parallel architecture and massive-thread working mechanism of modern GPUs,we build a performance model on a four-level stream processing pipeline,including data input,computation,output and transfer. We further conclude two key factors for GPGPU applications to achieve high performance:high intensity in both computation and memory access. To demonstrate the effectiveness of our model,we highlight its performance on three typical image processing applications,including Gaussian convolution,DCT and motion estimation. Guided by our performance model,some techniques are put forward to remove the performance bottleneck and achieve 4～10 times performance improvement. In practice,the experimental results fit well with the estimations of our proposed model.