It is important to note that the algorithm has a high degree of data
locality. All operations on the volume can be carried out in parallel
and at all times the data needed for an particlular operation is no
more than one grid step away. With these properties it runs
efficiently on a wide range of parallel and vector architectures. It
also scales well for both the size and accuracy of the problem.
Although not yet implemented the calculation can be made wavelength
dependent simply by performing the same operations (in parallel, space
permitting) on another 
with different scattering coefficients.
This promises splendid sunrise and sunset effects.
The second view dependent stage is less satisfactory from a parallel point of view in so far as the problem was parallelised in image space, replicating the entire scene database on each of the nodes of the machine and assigning each a different part of the image. The work of Corrie and Mackerras on caching schemes for volume rendering would be applicable to this problem [4].