You are partially right, I was fast and sloppy and I gave the impression all is jitted and it is not the reality.

The part of the translation is fine. However there are parts that are compiled beforehand (shaders for example and I can recall something about arm or other architectures, not sure now). And this is a crucial point of the extra performance, because some parts can be ported to more updated/efficient implementations, not because there are less services in the background.