Another trick is in stead of tracing multiple bounces per ray at once, to compute one indirect bounce every n-th frame using intermediate results from the previous (not the current) frame and then spatio-temporally smooth the results, possibly with on-screen bilateral filtering of the past few frames. I think one can see they do this when they move the light source. It takes a while for the indirect bounces to fade out. Perhaps they use screen space ray tracing for this (meaning no indirect bounces for occluded geometry).
reply