[tomooot]

Even modern high end graphics cards use abstractions of the base data to create vast amounts of the final output's fine detail. For example tessellation and other techniques used for complex geometry like compound curves, which allow millions or billions of polygons can be visually simulated without needing to be present as polygon data, increasing opportunity for processing parallelization, while reducing load on communication busses and VRAM.

As an example, you could probably represent something like the grip of this FLIR camera in a couple hundred polygons and surface/curve definitions to help the rendering engine tesselate correctly. On the other hand, this overall scan is 357000 vertexes. Sure you can simplify it and bake a bunch of the texture into a normal map, but that then requires manually reworking the texture map and various other postprocessing steps to avoid creating a glitchy mess.

https://i.imgur.com/aAwoiXU.png