I depend heavily upon the capabilities of "real-time" GPU hardware, not only because I really don't have the time to wait, but also because I know I shouldn't have to.
In the environment in which I work, Blender, there is already a GPU-based "preview render" feature which cranks-out what's on the screen using the GPU (GLSL) hardware. You'd be quite surprised just how far this alone can get you, if you put your mind to it.
Real-world rendering tasks always use "nodes," and I have likewise found that many of Blender's rendering and texturing features do exploit the GPU ... partly because Blender has a built-in game engine. But I am also very actively exploring its PyNodes (Python nodes) features to see just how much can be reached through programming.
I agree that "it is the software that's lacking now, not the hardware."
At the same time, I do not frame this requirement in terms of "real time," because really it is not a "real time" requirement, and that's what makes it so different from "gaming." This requirement is really for hardware-accelerated rendering nodes. You are ... first of all ... using the node-based rendering capabilities of your software (not "the big fat Render button").
You just can't expect to support "the big fat instant-gratification Render Button."
You are doing work without the expectation of "real time." Instead, you are asking the software to, at particular times and in particular ways, hand-off tasks to the GPU for completion. You are consciously designing both the render-network and the project itself to exploit hardware capabilities.
You're not shooting for "frames per second," but rather, "seconds per frame."
I can see that Blender can do a lot of this now, and as I further explore the Python features (it sure helps to be an experienced programmer! ) I'm sure that I have only scratched the surface.
Anyhow... it definitely qualifies as "a huge win." I have shaved numerous days off project times with even the little bit that I have done. The results are "satisfactory to my requirements" (which may or may not be yours!) and the speed increase is compelling.
If folks are using GPUs for ordinary array-math tasks, then I'm sure they'll be using them more and more as a "vector processor" for otherwise-conventional rendering pipelines.