Why does the graphics pipeline specify that the depth test is performed after fragment processing, if that results in a lot of wasted work? When is it better to use the process on the left?

Reason to sort: prefer front to back because otherwise early z out doesn't really happen

I'm still a little confused about this method. If we sort the geometry first, don't we still need to perform pixel-level depth tests at the Frame-Buffer Ops step in case of geometry intersection? If this is the case, where does the optimization come in?