I've seen this before and read all their papers, and can't really say that I find it that interesting. First off, I think people have this confused correlation between raytracing and soft shadows, diffuse interreflection, caustics and the like. Just because the card produces pixels without rasterization doesn't mean that any of the effects that we associate with pretty offline raytraced images are feasible, or even possible. Getting decently GI takes much much much more work than what they are doing here.
There isn't really anything in those images that you can't currently accomplish with modern graphics hardware besides the reflections between objects. Hard shadows? Check. Crappy, low sampled "soft" shadows? Check. Sure, the reflections are nice, but honestly do we really need that so badly? How often in games or in real life do you see massive numbers of highly specular objects in proximity to each other. The biggest problem is still transparency, but there are ways to achieve the right result.
You could add back the environment mapping and lightmaps to improve the visual quality, but the aliasing is simply unacceptable for modern real-time graphics. Also, you would have to add a lot more circuitry to handle anything near the complexity of current shaders.
I think people continue to overlook the difference in how zbuffered scanline rasterization and raytracing operate. In their naive form, both would be linear in runtime with the number of objects to be rendered. Raytracing makes the claim that it can operate in logarithmic time with spacial acceleration, while rasterization would be linear. While this may be true if you tried to render your scene in a single Draw() call, rasterization can benefit greatly from spatial subdivision as well and can use recent hardware advances like top of the pipe Z reject where you can render your scene once very quickly as a first step and then only compute shading once per pixel.
The biggest difference is in data coherency. When raytracing, the entire scene must be available because without prior knowledge of the object's surface, a reflected ray could go anywhere. You need to have all of the scene data available. Often adjacent rays lead to vastly different reflected rays, and have very poor coherency of access. Rasterization differs where each object is independent of the scene. You could theoretically page geometry on and off the GPU for each draw and have no effect. With computational power outpacing memory access in the fashion it is, coherent access is needed to hide the latency of the memory system. You take the cache hit on the first read and get the next 7 for free or whatever. Clever use of rasterization hardware can decompose many of the portions of simple raytracing that are not memory coherent into rasterization steps that are.
I'm not saying that hardware raytracing won't be possible, but my guess is that it'll take at least another 10 years to be useful. And even when the computational power exists to support it, I feel that rasterization hardware would make much better use of it. For the majority of things, I feel that rasterization will be the way to go, and future GPUs will be general enough to let you implement raytracing, etc... for the things that need, but avoid the cost for the things that don't.
