Volume rendering on a distributed memory parallel computer

A prototype implementation of a splatting volume renderer (SVR) on a commercially available distributed memory MIMD (multiple instruction stream, multiple data stream) parallel processor, the nCUBE2, is described. Some relatively good rendering times can be achieved with the nCUBE SVR. Message-passing bottlenecks occur when large numbers of floating-point values have to be collected from every processor for every picture. For large images this is a severe limitation. An initial implementation of a SVR on a distributed memory parallel computer demonstrates the need for parallel computers with high-bandwidth connections between processors, and also for new parallelizable volume rendering algorithms. >