Depends what you want to do. There are several realtime raytracers on GPUs, we have one here at Stanford that was shown in the ATI booth at SIGGRAPH. The main issue is efficiency compared to other designs, like the Saarland RPU work which is pretty neat.
The rough estimate from the last raytracing talk I heard at SIGGRAPH was that we need to be on the order of 500Mray/s to be interesting. The best raytracers currently published use Cell and are ~100Mray/s for primary rays. When you start talking about global illumination and dynamic scenes, the requirements start to go up. However, few people do full shading of scenes like modern game engines do, so the numbers are generally ultra simple shading. I still think we are a good couple of generations of processor design until we can really think about doing fully raytraced games. But, I'd expect to continue to see hybrid designs come out and blur the line, using rasterization for primary hits and raytracing for secondary effects. For example, POM (parallax occlusion mapping), already does localized raytracing.
