The real-time visualization of large-scale terrain is a very complex and challenging problem, as the increasing resolution of geometry and texture has greatly exceeded the rendering capabilities of the available hardware. This is an active research area, and has attracted the interest of researchers for decades. To solve this problem, LOD-based algorithms of various forms have been proposed. These aim to reduce the complexity of the geometry and decrease the number of triangles that must be rendered. In this paper we will mainly focus on LOD generation, data scheduling and terrain popping, and put forward a new approach for terrain rendering. To promote rendering efficiency, we will take advantage of the programmability of GPU and immigrate some algorithms from CPU to GPU.