The gp comment speaks more of freight than of passengers. But the same analogy applies… trains are great when you need to transport a whole bunch of people exactly the same way, but they’re less useful if every person needs to start and end at a totally different place.

In the same sense, parallelization (and computing on a GPU) are great when you need to perform the same computation on a huge number of data. When each computation is more independent, it might make more sense to perform it through the CPU.