The new generation of commodity graphics cards with significant on-board video memory has become widely popular and provides high-performance rendering and flexibility. One of the features to be exploited is the use of the on-board video memory to store geometry information. This strategy significantly reduces the data transfer overhead from sending geometry data over the (AGP) bus interface from main memory to the graphics card. However, taking advantage of cached geometry is not a trivial task because the data models often exceed the memory size of the graphics card. In this paper we present a dynamic cached geometry manager (CGM) to address this issue. We show how this technique improves the performance of real-time view-dependent level-of-detail (LOD) selection and rendering algorithms of large data sets. Our approach has been analyzed over two view-dependent progressive mesh (VDPM) frameworks: one for rendering of arbitrary manifold 3D meshes, and one for terrain visualization.