Oh yeah, my last post made me think of the Asimov story boris posted -
The Last Question.
I think the growth is probably stable if you use the right scale. For instance, if you use orders of magnitude then it goes 1 hour, 10 hours, 100 hours, 1000 hours, 10000 hours. So it's only five "steps" really. Whereas you have many "steps" under 1 hour (down to tiny fractions of a second) and over 1 hour (up to trillions of years). So there's an illusion of "clustering" if your scale is ordinary (I don't know the terminology, but if each interval is n instead of 10^n). Though I think the growth is still very slow after the 1-year point. The difference between a 1,000-year program and a 1,000,000-year program isn't trivial, right?
Another problem is that supercomputers have already solved many of the mid-range problems. Stuff like "find the first n primes" has also gotten out of hand, because distributed computing projects have have already found the answer for a large n.
I do think rendering typically falls in the middle range, and it seems to fit the criteria.