Where is FGRP5 computing power going to?

No.  The app hasn't changed so the precision of calculations hasn't changed.

It has to do with the data (LATeah00xxf.dat) where the value of xx changes or with the 'template' files you receive with each task, which include values for certain parameters which must control how many times the calculations 'loop' or how long a loop might take to complete.  To get some idea, take a look at what gets returned to the project which you can easily do by clicking on the task name on the website for any returned task of interest.  Below are a couple of examples.

Please note:  Outside of the scientists and admin staff who run the project, the volunteers (like me) who try to answer questions here, are just making assumptions based on what they have observed over time.  In the absence of 'official' statements, you should make your own observations and form your own opinions.

Excerpt from the Stderr Output for a LATeah0019F Task

% Sky point 8/8
% Starting semicoherent search over f0 and f1.
% nf1dots: 544 df1dot: 1.843348696e-15 f1dot_start: -7.5e-11 f1dot_band: 1e-12
% Filling array of photon pairs
................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
% C 8 0
% Time spent on semicoherent stage: 34120.7665s

This is right near the end of the stderr output, at the end of the main calculations stage for a 19F task.  The last 'sky point' (#8 out of 8) had been processed and the checkpoint for that sky point (% C 8 0) had been written.  The total time spent on all 8 sky points is given.  I didn't count them, but I assume the long list of 'dots' is equal to nf1dots (the number of f1 dots - my interpretation) which is an indication of the number of major loops in the calculations.  I also assume that 8 times 544 (4352) is the total number of these major loops.

Excerpt from the Stderr Output for a LATeah0018F Task

% Sky point 81/81
% Starting semicoherent search over f0 and f1.
% nf1dots: 55 df1dot: 1.855951967e-15 f1dot_start: -1e-13 f1dot_band: 1e-13
% Filling array of photon pairs
.......................................................
% C 81 0
% Time spent on semicoherent stage: 21378.7005s

Note the big difference for an 18F task - 81 sky points but only 55 f1dots and 55 '.' symbols (I did count them this time :-) ).  So, in total, (if what I already stated is correct) there are 81 times 55 (4455) major loops, much the same as last time.  I assume f0 and f1 are frequencies, and the values for df1dot, f1dot_start, and f1dot_band are numbers that come from the 'templates' that are downloaded with a task and that these values have an impact on how long a 'loop' takes.

Why don't you have a browse through the contents of the template files  for particular tasks and correlate the numbers there with the parameters mentioned above.  Maybe you will be able to figure out why one set of 'major loops' takes a lot longer than the other.  You could then come back and let us all know the answer :-).