A method is provided for distributed element rendering with particular application for feature films and other demanding content creation using scenes of high complexity requiring global illumination. A persistent centralized scheduler receives shading queries that are added to a request queue, determines an assignment of the request queue to hardware resources based on a resource map, and processes the request queue according to the assignment to render frames of one or more scenes. The resource map may be built by the scheduler and indicates local scene geometry cached for each of the hardware resources. By generating a full set of camera rays at each hardware resource, global illumination shading and other rendering can proceed independently against local geometry caches for high parallelism. Redundant computations are reduced through the scheduler, which may cache frequently requested rendering results including tessellation, shading, and level of detail.