Ok - so the "afaik" at the start was a good idea
A quick search only brought up comparisions to normal mapping - which does create additional geometry.
Though todays GPUs must be a lot better in raytraycing-processes, if this approach is the faster one. Or I'm just to stupid, to see any other approach doing these shadows, then "by pixel" as well. And "number of pixels" by "number of light sources" should give a huge number of calculations.

Anyway, the paragraph about tesslations holds on

That was a 'Nerd Moment' with ruyven.

Probably not. As well as "normal" "normal mapping", parallax occlusion mapping is, afaik, just a very convenient format to transport detailed information to the graphic card.
But as soon, as it is applied, all the little numbs and dents have to be calculated as actual geometry, to enable further effects (like dynamic and multiple (self) shadowing). So let's say a stone wall looks a lot better, if you use p-o-mapping instead of bump mapping. But while the bump-mapped wall still consists of only two triangles and a couple of image layers, the p-o-mapped version of the wall uses a seperate polygon for each side of each brick and each surface of each crack in this wall, thereby increasing gpu-load by several orders of magnitude.

The same applies for tesselation - it's nice, if you tesselate a curve into 125153 bits, until every single pixel has represents its own polygon and therefore the curve looks perfectly smooth. And tesselation allows you, to do this, without loading the cpu.
But you still apply the full load of such detailed geometry to the gpu - and frankly: In a 3 GHz-Quadcore-GTX275 System running e.g. Crysis, it's not exactly the CPU, thats overloaded and the GPU, that's idling around...
So personally, I'm quite happy, that developers don't use these techniques at the moment.