A one-pass rendering scheme of 3D dynamic scenes using modern graphics processing units (GPUs) and graphical interfaces is considered. This scheme uses the following methods and techniques: frustum culling, hardware occlusion queries, fragmentation, and caching of command buffers. These methods require significant computational resources, and the amount of work on the stages of graphics pipeline depends on their results. For efficient rendering, the balanced use of resources during the pipelined processing and transferring graphics data is important. A performance model of the graphics pipeline as applied to rendering 3D dynamic scenes that makes it possible to estimate the required resources, depending on the selected base methods and characteristics of the scene being rendered is proposed. In distinction from the existing methods and models, the proposed model allows one to calculate the cost of composing command buffers using various recording techniques, calculate the cost of transferring, executing, and loading the results of hardware occlusion queries. Formulas for calculating the frame rendering time depending on the number of occlusion queries are derived. A method for estimating the number of hardware occlusion queries needed for efficient rendering of dynamic scenes is proposed. Computational experiments demonstrating the relevance of the proposed model and the efficiency of the developed method for rendering large dynamic scenes are carried out.