It is often the case that numerical algorithms suitable for scalar architectures cannot be easily adapted for use on massively parallel architectures. This is definitely the case for many of the well known computer graphics which rely on ``divide and conquer'' strategies (eg octree and other spatial bounding methods). Although highly efficient on machines with shared memory (before scalability becomes a problem) they do not translate easily to data parallel machines like the Thinking Machines CM2/5's.
This paper addresses the problem of global illumination in a data
parallel fashion. The particular problem is that of the simulation of
the multiple scattering of light by diffuse atmospheric features like
clouds, air borne dust particles etc. This problem is very
computationally expensive using traditional methods such as
Monte-Carlo/distributed ray tracing. It is also an excellent problem
to explore since it is one of the least sensitive to rough
discretization in direction.