Those are not 100% decoupled from one another. Just as you alluded to lowering the polygon count to a level to keep most of the quality with lower requirements just cranking up the polygons without limit doesn't necessarily increase the quality of the render. It just just makes it take longer.
If have 10 polygons per pixel on the screen who is going to see them? They all can't be presented on the screen. Sure some might be clipped from view but 9 for every pixel?
You tell me how you get 1.5 billion polygons on a 800x600 screen.
If no one can see them what impact does it have on the final result? Completely obscured from just about all angles polygons aren't going to contribute much.
Can play the game of "I need 100 million polygons"... "too small? then need 1 Billion" ... "too small? then need 10 billion" ... All the while the human eye has constant constraints on what it can see.
Right, because it mostly works for sizable group of people now.
I'm sure there are some folks rendering for 4K video, IMAX, or tiny few for 8K prototype system. 12GB may not necessarily be the path as much as more uniform access to main RAM and better ( coherence with RAM and VRAM ) , bigger (capacity) L3 on GPGPU cards.
Cards will probably hit way points at 8 and 10 before getting anywhere close to 12. Could get to point where VRAM is just treated as a humongous L4 cache of main RAM. That would mean need enough to work on the current computational fronts/waves and just page in as necesary a complete the whole picture.
Maybe you should stop pretending that you know it all and read what other people are writing instead of trying to interpret what they wrote.
When you design in 3D you aren't doing it at a set resolution, instead you are doing it at a scale. As Flat Five posted just above me, all object in your scene is built individually and scaled in relation to the overall scene. ALL THE DETAIL OF EACH OBJECT STILL EXIST AND IS INCLUDED IN THE SCENE. This is how 3D modelling in blender works.
i'm presently working on a T-72 model tank for an animation project of mine. I started up by modelling one track wheel with all the detail like bolt and such. I then designed the piece on which the wheels are joined etc, etc... You can scale down to see the hexagonal bolt heads... And this is important, because in my animation the camera will zoom in on the lower side of the tank and the wheels will be visible. And when it comes time to render then all of the quads that exist in this model has to exist and need to be processed to see what impact it has on texture, light and shadow and it has to do thise for every frame of animation taking into account those that came before it also. This eats up vram or plain ram fast. Why do you think Pixar and others have humongous render farms... But I know of single artist and small studios who are near to break that 6Gb barrier or have adapted their workflow around it.
We're not discussing raster based 2D graphics here. We aren't talking about low poly game asset either.