The performance measures for the algorithm were done using a 256x256x113 CT dataset of a human skull. The data was scaled by a factor of 2 in the Z-direction and the dimensions for the generated image were 375x375 pixels.
The speedup was almost linear when adaptive ray tracing was not used. If adaptive ray tracing was not used the speedup with 32 nodes was 28. When it was used, the speedup for the same number of nodes dropped to 23. An interesting fact is that, although the speedup was lower the rendering time was faster. This clearly shows that speedup is not a good performance measure.
Another interesting result was the measurements done using MTOOL [32] that indicate only 17% of the time is spent waiting for memory in a one processor configuration. In some sense this number indicates how much of ray casting time is spent doing useful computations.