I remember there was some kind of controversy regarding Nvidia not having the architecture needed for doing "proper" async compute, and this theory was somewhat vigorously used to prove that AMD was the only future ahead. I understand the basic principles, but I'm not knowledgeable enough to understand if that chart there beyond first glance accurately tells another story or not. Anyone?
The problem with Nvidia's async compute on Maxwell comes from a variety of issues: latency, scheduler, etc. There was even a test done on b3d which tested the scheduler, and showed that async compute was extremely fast provided that the scheduler was not filled to the brim. Changing the scheduler and fixing the timing/latency issues in Pascal will go a long way to improve the async compute performance compared to Maxwell.
Of course, we haven't seen anyone post any in-depth tests on the Pascal scheduler, as well as other architecture-level tests, so this is still speculation for the time being.