I don't quite understand the graphs. If all we did was move rasterization to a separate thread, how does it take any time in the frame at all? Are we measuring instead the cost of synchronization/contention? If we measured rasterization time exactly the same, wouldn't the absolute quantities be relatively similar since we aren't changing the rasterization algorithm itself?

One interesting thing I realized reading this is that I've never bothered looking at Direct2D, having used Direct3D forever.