mihen :
Dedicating hardware just to raytracing is dumb for a consumer card right now. Waste of resources for something that will probably be do-able by the same compute cores the rest of the GPU is using.
Nvidia actually supports this on their Pascal GPUs. According to their performance numbers a GTX 1080 Ti is only about 1/10th as fast as their RTX 2080 Ti (measured in gigarays/sec). That's the power of fixed-function hardware.
I do agree that it would be nice to find some kind of balance, like their Tensor cores, where the solution is just generic enough that you could possibly use it to accelerate structurally similar algorithms. Of course, that might already be the case but they just haven't yet publicized it (perhaps due to SDK support needing more time to mature).
mihen :
When we do get real-time raytracing, it will probably ... shoot a ray out for each pixel and lets it bounce to a light source if one exists.
This provides only
some of the benefits available with Nvidia's current technology. Some of the most powerful and difficult to emulate effects supported by ray tracing are lighting-related. Indirect lighting and caustics, in particular. Also, soft shadows are easier to implement in ways that play well with other rendering techniques. Even motion blur and depth of field would require more than one ray per pixel.
About the only things you can get with one ray per pixel that are any better than traditional rasterizers are accurate reflections and refraction. However, as both are quite susceptible to aliasing, even then you'd want to supersample.