Algorithm for ray-AABB test behind the acceleration structures

Does OptiX use robust ray-AABB intersection test in the acceleration structures?

UPD:

To be more specific, I get different results for my launches, depending on the coordinates of the vertices in the triangle mesh and the position of the ray source. Namely, the result is almost unusable if the geometry scaled 100 times up.

AFAIK, there are three sources of precision loss in my rendered.

  1. My custom ray-triangle intersection possesses some precision losses.
  2. In recursive ray casting I change the ray source to the point of current intersection, which was found with some precision losses.
  3. Ray-AABB intersection test in the BVH traversal.

While I have full control over the first two items, the third one seems to be hidden. And I need to know that to estimate the effects of scene geometry scaling and to find the acceptable range.

For example, how does the algorithm relate to the described here http://jcgt.org/published/0002/02/02/paper.pdf?