Some GPU based methods are faster when they rely on raw parallel processing power, even if it caused a huge amount of work to be discarded. These chips are well suited for regular, embarassingly parallel, dumb work. This starts to drop sharply with increasing branching and the amount of unordered memory accesses, although GPUs get better and better at that. So it can be better right now to do the dumb thing instead of the smart thing.

I think that sorting on the GPU is a good example: a parallel quicksort is hard to implement and maps badly to GPU threads. A bitonic sort algorithm is likely to be much faster, although it is a very dumb algorithm.