In this paper we present our research efforts towards a scalable volume rendering architecture for the real-time visualization of dynamically changing high-resolution datasets. Using a linearly skewed memory interleaving we were able to develop a parallel dataflow model that leads to local, fixed-bandwidth interconnections between processing elements. This parallel dataflow model differs from previous work in that it requires no global communication of data except at the pixel level. Using this dataflow model we are developing Cube-4, an architecture that is scalable to very high performances and allows for modular and extensible hardware implementations.