I don't have any hard facts for you, but I assume that they are just looping over all of the triangles in the mesh and doing a ray/tri intersect on it. I assumed this because it is rather slow for a large mesh (I used it for a while on some large terrain meshes). To speed it up, I wrote a KD-tree to divide up the triangles into smaller chunks and traverse that tree with the ray before looping over the tris that are in a small bounding box (leaf of the tree).

I can't think of how they would accelerate the intersect method unless they do some kind of hierarchy or grid to reduce the set of triangles to test with.

I haven't done any timings or tried to look at the code, so I am not absolutely certain, but I know my kd-tree is faster.