To be honest, although GPUs are super cool, I have to admit a bit of disappointment in practice. I don't know about databases, but in the machine learning community ones sees usually about 6x to 20x speedup on GPU. This is amazing, of course, but far below what, way back, naively thought we'd be able to get from GPUs -- i.e., 1000 processors instead of 1, we should get a 1000x speed-up, right? Not so at all. So many architectural issues (i/o, memory, even sometimes algorithmic limitations) get in the way of achieve true parallel linear performance.