This is the first article I've seen that explains well how GPUs can/are being used for practical applications along with what can be achieved and some of the issues.
GPU's have been used for all sorts of "practical" computations for half a decade now, but the really interesting part here is that CFD has been particularly GPU-resistant using existing algorithms. See the Xeon Phi processor, etc. for non-GPU approaches to throwing dedicated hardware at the problem. It's easy to underestimate the enormity of this quote, but "starting from scratch" when necessary is something SpaceX excels at:
I am grateful to SpaceX for allowing us to basically start from scratch on CFD and in many ways reinventing the wheel.
It's hard to gain sufficient insight from TFA but it sounds like this is as big as hidden-line-removal in computer graphics and that they've moved CFD to the boundary conditions and made that GPU-computable, which is like solving two or three orders of magnitude at once.