The test model (18 polygons) was a white room with a square light suspended from the ceiling and a colored cube on the floor. For an image of size 512x512, brute-force path-tracing with 40 paths per pixel and 5 generations per path took one hour on an HP700 workstation and cast 100 million rays altogether, including those biased towards the light source. No variance-reduction techniques such as hierarchical sampling were used; these would have greatly reduced the noise at the cost of an increased rendering time.
My algorithm took twice as long on the same model and workstation and at the same resolution, but shot one-fourth as many rays. If the scene were more complicated, then the fewer rays shot by my algorithm would reduce the gap in the running times. An average of one percent of the sphere cells had to be traced at each pixel. (Notice that the cache-hit ratio increases with larger images since screen pixels have more locality in first-generation surface intersections.)
A more complex scene (161 polygons) of a large hall with a table and two sofas, six light sources (four windows and two spot lights), took only a little longer to render (three hours) and the results of diffuse-diffuse interactions were clearly visible on some walls and pillars, under the table, and in dimly lit areas. The noise in bright areas could have been reduced by using a variance-reducing scheme when randomly sampling the light sources.