This is expected to give memory saving ob GPU, hopefully without
Comes quite directly from the following paper:
Current code is slower, most likely because it lacs closes child
traversal heuristic (other than that it's expected to be same
amount of ray-to-AABB intersections). I'll look into solving this
regression if someone confirms any measurable GPU memory improvement
because for now memory usage on GTX760 and C2075 stays the same
before and after the patch.