In this talk Tim Sweeney discusses the long term future of computing and its impact on society, and specifically that of 3D graphics technology. It’s something I’ve asked myself before, how far away are photorealistic realtime 3D graphics? When will games and computer generated graphics in movies be realistic to the point of indistinguishable from the real thing?
The route that Sweeney follows to answer this question is to first ask what the limits of our perception are, in particular our vision (eyes plus visual cortex). Any improvements beyond these limits cannot be perceived, so they establish a definitive upper bound on realism. And this bound can be convincingly determined due to well established knowledge of our biology and sensory apparatus. A few numbers are given, for example in terms of maximum resolution and frames per second that we can perceive.
These numbers together with somewhat less objective intuitions as to which graphics approximations (such as lighting algorithms) are necessary to achieve realism, yield a factor of 2000 times today’s hardware performance to reach the target.
Sweeney then rightly notes that some aspects of presenting realistic graphics are not just a matter of simulating physical processes that are completely understood, but rather require theoretical advances. The main one is artificial intelligence, necessary to realistically simulate characters in terms of animation and even behavior. Still, the term photorealistic doesn’t necessarily include things like these.
Summing up, I’d say the 2000x figure is his rough answer, and it seems to me about right. How does this translate into years? You could use Moore’s law, but I suspect that would underestimate the potential improvements in graphics performance which could advance faster thanks to the particular parallel nature of rendering. So, I’d venture circa 2020 to get a nice round figure.