Because the data exceeds the 32 bit addressing limit of the PC's in the cluster, we have to explicitly manage memory in a software distributed shared memory (DSM). The DSM uses fixed ownership of data augmented with caching of non local data. This image shows which node in the cluster owns which portions of the data.

Ray tracing is inherently parallel, and we use standard image space parallelism for parallel acceleration. This image shows which processors in the cluster rendered which tiles.

We render isosurfaces using the full resolution data at all times. The data is a regular grid of scalars, 2048x2048x1920 or 7.5 GB in size. This series of images demonstrates the complexity of isosurface that the high resolution data yields. The viewpoint zooms in starting from an initial vantage point that shows the overall structure of the isosurface, to a closeup view that shows a small piece of the surface.

A unique aspect of our ray tracing solution is that the isovalue can be changed interactively. Because we compute isosurface intersections on the fly, changes in isovalue incur no more computation than changes in viewpoint. This series of images shows different isosurfaces for a single vantage point.