One of the main problems of rendering volume datasets is the amount of memory used to hold the dataset alone. This caused architectural problems in both shared memory machines (memory contention bottleneck) and distributed memory machines (network congestion). SIMD machines try to eliminate both problems by distributing the memory very finely across a large number of well connected processors [16].
By using a central host computer, SIMD also minimize the cost of memory usage in program code, as only one copy of the code is broadcast from the host to all the processors. The algorithm described below was implemented on the Connection Machine CM-2 [26], but it could be implemented efficiently in any SIMD architecture such as the MasPar MP-1 [25]. The goal of this algorithm was interactive use and the user can (by a user interface) interact with the computer as the computations are done.