This is a powerful feature, but it seems of limited usefulness in light of Vulkan. It strikes me that everything they’ve described there is an answer to limitations imposed by OpenGL - but Vulkan essentially lifts those restrictions. Compute shaders are first-class citizens from v1.0, and they can feed data into any stage of the traditional vertex->tesselation->geometry->fragment pipeline, each stage of which is programmable.
Since this is only for Turing and newer hardware, this can’t be applied as a driver enhancement to older hardware that can’t support Vulkan.
Another downside is that it introduces yet another GPU-specific “optimization” that’ll tend to clutter code (or make games run only on nVidia hardware - which I’m sure nVidia wouldn’t mind a bit), at least until AMD pushes out similar functionality. Vulkan runs on pretty much any reasonably recent GPU, and similar GPU-agnostic optimizations/abstractions could probably be handled via a compute shader.
It’s a cool feature and certainly worth exploring, but it seems to me that this is likely to only be a significant benefit to OpenGL programmers developing for the newest nVidia GPUs, at least for the immediate future.