Based on what I've gathered in the b3d forums, there is an issue with terminology here. Both Maxwell and Kepler can do asynchronous dispatch of compute threads (what most, including Nvidia, would call async compute).
What they can't do, however, is concurrent async compute and graphics (what Oxide and AMD refer to, doing graphics and compute at the same time in parallel).
The picture that Silverforce11 posted
earlier is a decent explanation, see how the orange blocks are split from everything else in the top timeframe, and in the bottom they are mixed into the blue blocks, filling up resources that would have otherwise not been used? That is concurrent compute and graphics (at least as far as I can tell).
The picture that Carfax has posted shows that Maxwell definitely can do 32 compute threads in parallel. I'm not sure if it's a good demonstration of async compute, because those threads are the same length and have no reason to stall or pass each other.
What should be fairly easy to tell, however, is that they don't fill up any resources that are unused during the graphics rendering, because the offset between pure compute and compute+graphics is nearly the same no matter how much compute is loaded onto the GPU. That means either Maxwell has no unused resources during graphics rendering (unlikely) or compute is calculated before or after graphics in that program.