Volume rendering with texture mapping hardware is a fast volume rendering method available on high-end workstations. However, limited texture memory often prevents the method from being used to render large volume data efficiently. In this paper, we propose a new approach to fast rendering of large volume data with texture mapping hardware. Based on a new volume-loading pipeline, the volume data is preprocessed in such a way that only the volume data that contains object voxels are loaded into texture memory and resampled for rendering. Moreover, if classification threshold is changed, our algorithm classifies and processes the raw volume data accordingly nearly in real time. Our tests show that about 40% to 60% rendering time is saved in our method for large volume data.