I can definitely see value in offloading a bunch of number crunching on sufficiently large data sets to a graphics card (particularly parts that are trivially p

I can definitely see value in offloading a bunch of number crunching on sufficiently large data sets to a graphics card (particularly parts that are trivially parallelizable). But to me, that seems like an optimization a database would make in its query planner, not one that you'd necessarily want to build an entire standalone product around.