2000x GPU Performance Needed To Reach Anatomical Graphics Limits For Gaming?
Vigile writes "In a talk earlier this year at DICE, Epic Games' Tim Sweeney discussed the state of computing hardware as it relates to gaming. While there is a rising sentiment in the gaming world that the current generation consoles are 'good enough' and that the next generation of consoles might be the last, Sweeney thinks that is way off base. He debates the claim with some interesting numbers, including the amount of processing and triangle power required to match human anatomical peaks. While we are only a factor of 50x from the necessary level of triangle processing, there is 2000x increase required to meet the 5000 TFLOPS Sweeney thinks will be needed for the 8000x4000 resolution screens of the future. It would seem that the 'good enough' sentiment is still a long way off for developers."
My question is this: how much more will games have to cost to support the development to this level of detail?
The soylentnews experiment has been a dismal failure.
I think 2000x GPU power is very much underestimating the potential for a number of reasons:
1: Raytracing / global illumination. In comparison to games with true global illumination, current technology 3D worlds with only direct illumination (or scanline rendering) look crude and unconvincing. Objects appear 'cookie-cutter' like and colours tend not to gel with the overall 3D landscape.
Toy Story 3 took around 7 hours to render each frame. To render in real-time for a video game (say 60 FPS), you would need a processor that was around 1 million times faster than what we have today. And AFAIK, that's mostly using Reyes rendering (which incorporates mostly rasterization techniques with only minimal ray tracing.
2: Worlds made of atoms, voxels or points. This makes a world of difference for both the user and the designer. Walls can be broken through realistically, water can flow properly, and explosions will eat away at the scenery.
2000x? Pah, try 2 TRILLION as a starting point.
Why OpalCalc is the best Windows calc
To me the real problem is focusing on the wrong details. Take Skyrim for example. Is it really a big thing if they, say, tripled the detail on the existing characters? Do the NPCs need pores or drops of sweat?
Or would it be more interesting to walk into Whiterun, and there's a 100 NPCs walking around, or you assault a fort with the Stormcloaks and there's 100 other soldiers at your side attacking the 100 Imperials in the fort, and clouds of arrows raining down [nice knowing ya, shieldless dual wielders :-) ]? It's a "more detailed objects" versus "more objects in the world" sort of argument, I guess. I'd rather see the power applied to "more objects" at this point, IMHO.
We passed a million polygons on screen over a decade ago. Your telephone can just about do that today (the iPhone 4S does 30 million per second), modern game consoles that came out 7 years ago will do about ten times that (500 million per second on a 360), and a modern high-end PC probably does ten times that again.
In other words, we're at the point where we're using rasterization to push 100 million polygons, and raytracing is still so much slower that it's not even remotely practical to duplicate the same quality. Intel's latest attempts to do so have produced low-resolution low-quality results that still require a massive array of hardware. They're basically throwing eight PCs worth of hardware at the problem. About all the demos do is demonstrate that it's easier to calculate accurate reflection and refraction with raytracing.
In other words, you either mis-remembered Intel's estimate, or their estimate was laughably inaccurate.