So far I have been holding my tongue on posting an example where the layout level portal filter is advantageous for performance relative to just using the relationship to filter. I was reluctant to post it, because in general, I completely agree with the sentiments already expressed here, and I did not want to detract from that. But, seeing so many final-looking comments, I figured it probably would not hurt to call out one crumb of exception.
Just as with so many things FileMaker, it involves knowing some of the quirks of the product, and how they can work for or against you.
Here's the scenario where the layout portal filters provided better performance:
I once was working on a system where they asked me to have a look at a portal that was part of a log viewer. The viewer had various filter features, and all such filters were implemented at the relationship level. They had found that the log viewer did not scale -- it was fine at first, but once a large number of log records had been accumulated, the viewer was unusable when its date range filter feature was applied.
The problem was that the relationships used inequality operators ( >, ≥, etc.) , and at that time (v.11), such relationships against a very tall table were dog slow.
The solution was to split the filtering implementation up, and hand off each part to the part of FileMaker that could handle it most efficiently. I reworked the filtering to leave all of the exact-matching filter aspects implemented at the relationship level. Those narrowed down the set of matched records to a large degree. The date range inequality was handled at the layout portal filter level. This arrangement had perfectly acceptable performance, despite having to send records down to the client which would later be filtered out -- simply because we got out of the business of using the ≥ and ≤ in the relationship. There was less penalty to send that data across the network than there was evaluating the inequality relationship.
The above situation was back in v.11. Many things have certainly changed since v.11, but if confronted with the same problem today, I wouldn't rule out trying a split filter approach to see what happens. (In general, I do my best to avoid inequalities in relationships, so I am not even aware as to whether this is still a performance pain point, or if it has improved with time.)
Hope this is an enjoyable crumb of an exception.