Is Möller-Trumbore ray intersection the fastest?

There is a paper from 2016 where the authors claim

Running under ideal experimental conditions, our algorithm is always faster than the standard Möller and Trumbore algorithm, and faster than a highly tuned modern version of it except at very high ray-triangle hit rates.

Source: Doug Baldwin and Michael Weber, Fast Ray-Triangle Intersections by Coordinate Transformation, Journal of Computer Graphics Techniques (JCGT), vol. 5, no. 3, 39-49, 2016

Available online http://jcgt.org/published/0005/03/03/

Be cautious of the Weber algorithm. While it may be faster, I'm seeing a good amount of intersections falsely identified as not intersecting. The paper states:

This series of calculations can terminate early if t is too small or large to represent a valid intersection, or if b1 is out of the range that permits an intersection.

I've seen about 2-3% of my mesh fail early because 't' was too small. I'm still troubleshooting, but it looks like the inverse of P is causing my rotated direction vector to be too large, equating to a small 't'.

On the other hand, you can also get false intersections with the MT algorithm if epsilon isn't set correctly.