So why implement this all in Fortran?

Long story short: speed.

Like I said in my previous description of unums, a big part of my interest in doing this implementation is to experiment with the number type myself. If the language is too slow, then the number of experiments I can run goes down and the overall usefulness of the code I write diminishes. Fortran is up there with C in terms of speed, and it's interoperable from a number of other languages such as C, Python, Julia, Matlab, and others. In other words, as long as I implement it efficiently, it should be fast and portable, two great qualities.

There are other reasons as well. Most obviously, I'm trying to refine my Fortran skills, and so this is a great project for that. Another is that being able to parallelize calculations done with unums seems to be a big selling point for Gustafson, and Fortran has several ways to parallelize its code. Granted, Fortran isn't unique in this regard, but it is one of the languages I feel is putting the most effort into parallelized numerical calculations. Perhaps this is another "speed" reason, but even just being able to test this ability seems more natural in Fortran than other languages.

I'll be very curious to see how much of a performance hit you take moving from native reals to unums, and how that varies depending on the task, I'll be thrilled if the result is that unums are "only" 5-10 times slower, which to me would be a fantastic result compared to data types that actually have dedicated hardware on the CPU to run quickly. 

Guess we'll find out soon, at least for basic tasks.