Abstract: The computational power required for direct volume rendering like ray-casting or volume ray-tracing can be provided by high-speed rendering architectures. However the increasing processor speed makes a performance bottleneck obvious - the volume memory. This paper describes a volume memory architecture that achieves at least a tenfold speed-up in read-out rate with moderate additional hardware. It has been simulated successfully. A multi-level cache system is used with software prefetching and latency hiding. Pre- and postcaches additionally speed up the read-out rate so that a 512³ data set stored in a single memory module can be rendered at 3.125 Hz.