In a wide range of scientific fields, 3D datasets production capabilities have widely evolved in recent years, especially with the rapid increase in their size. As a result, many large-scale applications, including visualization or processing, have become challenging to address. A solution to this issue lies in providing out-of-core algorithms specifically designed to handle datasets significantly larger than memory. In this article, we present a new approach that extends the broad interactive addressing principles already established in the field of out-of-core volume rendering on GPUs to allow on-demand processing during the visualization stage. We propose a pipeline designed to manage data as regular 3D grids regardless of the underlying application. It relies on a caching approach with a virtual memory addressing system coupled to an efficient parallel management on GPU to provide efficient access to data in interactive time. It allows any visualization or processing application to leverage the flexibility of its structure by managing multi-modality datasets. Furthermore, we show that our system delivers good performance on a single standard PC with low memory budget on the GPU.