Slashdot

It's no surprise that Intel is being bashed over their idea of real-time CPU ray-tracing. As anyone who has ever ray-traced will realize it's extremely slow. At times you're talking about HOURS PER FRAME while realistically you want at least 30 frames per second and even that isn't considered great by many gamers. It's going to take a HUGE and I mean HUGE increase in computation power before that happens. Rasterization techniques are tremendously faster and they look nearly as good as Ray-tracing for the most part. Considering that we're yet to reach a point in Rasterization where we don't need more processing power (Crysis in high resolution.) I don't see us moving away from it yet. The day when we declare that we have graphics cards more powerful than we need for Rasterization is when we start moving towards ray-tracing. That day isn't anytime soon unfortunately.

For years some claymation movies were set up by hand and shot frame by frame in a process called stop motion [wikipedia.org]. While adequate, the resulting film was typically unnatural and the movements very stiff compared to live actors.

Enter ILM and go motion [wikipedia.org]. Instead of filming static scenes, the clay was moved slightly during the shot to create a blurry frame. This blurry frame made the scene seem more realistic. The blur is what the eye picks up in the movie frame, so an actor walking in a scene is not a set of pinpoint focus shots but a series of blurs as the man moves.

Ray tracing is great for static scenes. But movement is the key to games that require this much detail, and so each frame should not be beautifully rendered framebuffers, but a mix of several framebuffers over the span of one frame. Star Wars did it great. Most computer games, not so much.

it's customers that drive the market, not developers. christ these guys sound like a bunch of OSS developers.

Is he the same developer who made a game (Crysis) so resource hungry that no gaming platform can handle it? Shouldn't we be asking someone who knows how to make a game look great on current hardware, such as Valve perhaps?

I ignored this story first time around because I assumed it must be an April fool's joke which I think is not unreasonable: Intel leading innovation in the GPU sector ....

Yeah, so it's going to take 3-5 years before anything real comes out of it. Do you think that process of using high-k hafnium in the 45 nm microprocessors was developed overnight? I'm sure intel is used to the R&D cycle, and 3-5 years is not unheard of. Besides, how much longer can you use rasterization "band aids" to address rending issues (reflections, shadows, light sources)? Rasterization is just a hack to try to implement features that simply "fall out" of ray tracing. Sure it's going to take computational power, but we're not going to be using pentium 75's.

Carmack didn't really bashed it, neither did Crytek. They just make it clear that you can't have rasterization on day N and have raytracing on day N+1. A 3-5 years transition period is very reasonable. Using raytracing optimally requires to change the whole data structure of the virtual world. It would require making new modeling tools, new rendering engines, integrating new possibilities into the game design.
Keep also in mind that Intel proposes this as a future way of doing rendering. Their hardware is not even here yet. Given this, any prediction below 3 years would be quite surprising.

Let's surmise for a minute:

The problem with ray tracing, as Carmack said, is that it will always be much slower than raster-based graphics with a given amount of computing power. He pointed out that there's nothing impressive about Intel's demo of a game from two generations ago running sort of acceptably at moderate resolution on an overpowered Intel demo system. He said that they'll never be able to produce a ray traced engine competitive with the state of the art raster-based games, so the ray tracing, while technically satisfying, will in every case offer poor performance for inferior graphics.

All of this boils down to a time lag. If raster graphics can do something in 2008, ray tracing can do it in 2012, etc. What if raster graphics stopped progressing for four years? Then ray tracing would have a chance to catch up, perhaps leading to new engines and APIs based on ray tracing, which would ensure long term use.

But wait...raster graphics have already been at a standstill for two years, for the first time since their inception. When the 360 came out and then the 8800 line showed up to put it firmly in its technical place, gaming graphics capabilities suddenly stopped. Not only did nVidia have its first unassailable lead over ATI in a long time, but suddenly the PC gaming market finally showed very strong signs of finally dying. Most of the remaining PC game developers shifted development to consoles, leading to (again as Carmack pointed out) a stationary graphical hardware target for new games. The overall number of PC gamers managed to stay high, but literally almost all of them were playing World of Warcraft, which has very low graphics card requirements.

Now two years have gone by, and WoW still dominates PC gaming, while only a few games have shown up that really push current hardware, with few people buying them. It's a pity that the most graphically impressive game is also quite mediocre when it comes to gameplay. There's very little market pressure on nVidia outside of the small enthusiast community, and they've managed to milk a 4x hardware lead over consoles for an unprecedented length of time. The graphics card industry used to beat the living crap out of Moore's Law, but now they've managed to eek out a 10% improvement in over two years, which is just sad. The next generation parts may or may not be coming soon, may or may not bring a large performance boost, and may or may not have any software at all to really justify their purchase.

Going waaaaay back to the beginning, CPU speeds over this same time period have been keeping up with their normal exponential increase in power. At this rate, it would only take two more generations of PC gaming failure for ray tracing on the CPU to catch up with rastering on the GPU, and if that happens, it could end up going to consoles. Hell, it might even be good for PC gaming's health. Currently most console players have a PC, but with its Intel integrated graphics it's only suited to playing games from 6-8 years ago. Already those same PCs can probably match that with ray tracing. If games were only dependent on CPU speed, they'd be a lot more accessible and easily played by a much larger part of the population.

Like all technology races, simplicity wins. If Intel provides tools that make it easier to develop ray tracing games, the GPU will be displaced.

But how much better do game graphics need to be?

I played the Crysis demo on a recent graphics card, and was suitably impressed for ten minutes. After that, it was the same old boring FPS that I stopped playing five years ago. Graphics seem stuck in the exponential curve of the uncanny valley, where incremental improvements in rendering add nothing to the image except to heighten that sense of 'almost there' that signals to the brain that it's *not* photorealism.

This isn't meant to be the same old "it's the gameplay, stupid" rant that we get here. It's simply to question why any real work is being done on rendering engines when we seem to long have passed the point of diminishing returns.

Personally, I'd love to see realtime raytracing see the light of day because I recognize the math behind it as more "pure" than rasterization. Of course there are several algorithmic hurdles preventing pure realtime raytracing from seeing the light of day, unless you start to hyperparallelize the operations in a dedicated GPU, and even then there are obstacles; in the worst cases, a ray can bounce along an infinite path, dividing into multiple segments as it goes, leading to infinitely branched recursion until some heuristic or another cuts it short. And as we all know, "heuristic" is a fancy word for "cheat".

Further, raytracing cannot handle advanced refraction and reflection effects, like the surface of water causing uneven illumination at the bottom of a pool, or a bright red ball casting a red spot on a white piece of paper, without preemptive "photon mapping", which is another cheat.

In short, we have not been able improve upon the original raytracing algorithms without "cheating reality". Modern raytracing that includes photon mapping is a hybrid anyway. So the raytracing purists really have nothing to stand on until there's enough hardware to accurately calculate the paths of quadrillions of photons at high resolution sixty times a second. I'm not saying we won't get there, I'm saying probably not within this decade.

The reality is, the only advantage raytracing has over rasterization is its ability to compute reflection, refraction, and some atmospheric effects (e.g. a spotlight or a laser causing a visible halo in its path) with "physical" accuracy. The capabilities of rasterization have grown leaps and bounds since the 1960s, roughly linearly in proportion to available hardware.

Purists be damned. A hybrid of each technique utilizing what it's good at (raytracing for reflection, refraction, and atmospheric halos, rasterization for drawing the physical objects, "photon mapping" for advanced reflection and refraction effects) is likely the best approach here.