What I see there is a lot more than just 3 listed types of work units for Amicable Numbers. I expect these numbers might be batch numbers rather than WU types (worth to find out), but those numbers don't seem to be in order. See also AN WU optimization.
These are some huge RAC numbers, lol. I have expected for 1080 Ti to reach ~1M RAC, 4M surprised me, but you haven't re-tested it. 6M for Titan (after re-test) seems reasonable.
Ah I didn't mean to imply types of work units like PPS-Sieve vs GFN on PrimeGrid, I guess the word I was looking for was batches. I think you're absolutely correct, and I'm going to update the article with that correction.
I do remember reading that article a while back, I guess 518 and 6141 in this article are examples of more GPU-optimised work units, whilst the 3148 is less so.
OK, I have 1st hand explanation for WU naming in the Amicable Project.
Example: amicable_10_20_20198_1519206302.310434_177
amicable_10_20 - app name
20198 - PID of the process that generated this WU
1519206302.310434 - timestamp
177 - number in the batch
Conclusion - It's only a coincidence than 518 WUs took much shorter time than 3148 WUs.