I need to run some more numbers, but the dismissal of raytracing on performance grounds seems inaccurate to me at first glance. Note that they inflate the number of ray intersections by including shadow tests. That number shrinks dramatically when you ignore shadowing and refraction as the final implementation does.