Single image super-resolution is a ill-posed problem which aims to characterize the texture pattern given a blurry and low-resolution image sample. Convolution neural network recently are introduced into super resolution to tackle this problem and further bringing forward progress in this field. Although state-of-the-art studies have obtain excellent performance by designing the structure and the way of connection in the convolution neural network, they ignore the use of high-order data to train more power model. In this paper, we propose a non-local second-order attention network for single image super resolution, which make the full use of the training data and further improve performance by non-local second-order attention. This attention scheme does not only provide a guideline to design the network, but also interpretable for super-resolution task. Extensive experiments and analyses have demonstrated our model exceed the state-of-the-arts models with similar parameters.