> I also wonder if you'd even bother writing code to do it on the GPU in the first place.

Some of your customers have large datasets that are only manageable with hardware acceleration, and others don't? Maybe you don't want to maintain two separate codebases?