HI Sascha,
That’s a great question! (I’m trying to simulate an LLM, simulating a human, here.)
The big difference is, that the Normal of the Distribute Points on Faces outputs the Normal of each Point, whereas the Normal, Captured from the Geometry, in the case of the Cube only has 6 Normals (one for each Face). And that happens to be significantly faster.
There may be situations, where using the Normal of the DPoF Node would be preferred, but I can’t think of any.
