Typically the cpu pre-renders all frames. It decodes and places every object according to the game code, then sends that info to the gpu. However many frames the cpu can pre-render in 1 second would be the fps limit. It's physically impossible for a gpu to finish rendering more frames than the cpu can send.
Ampere is changing the game with that. It's now possible for direct ram to gpu rendering, bypassing the cpu entirely, so the cpu takes on less importance for object placing, allowing it to concentrate on externals like physX and AI.
The biggest problem will be the SLI. It's effectively dead, buried, moot. A second card will not be of any help in DirectX 12 games other than the odd few setup for multi-gpu. A second card will only be of small help to the select few DirectX 11 games that have any decent SLI optimization. Which is pretty much nothing newer. With many of the older games, optimization for higher class cards was terrible, you end up getting better performance from a single card than what the SLI can accomplish.
The 3090 is a Titan Class card, pretty much not really designed for gaming, but for programming that takes serious advantage of mgpu and is best done with gpu hardware acceleration, not cpu. Like the stuff Disney or Pixar or LucasArts deals with.
It's an expensive waste of money, even nvidia understands that, which is why the 3090 is now the Only 3000 series card to offer multi-gang setups.
Oh, and 150w is a standard, not a reality. Reality is 60w per hot pin at 12v, 5A. An 8 can handle upto 180w, a 6 is good for upto 120w. Even the pcie x16 slot is good for 90+ watts. That's how you get cards like the Radeon 295x2 with a 450w power draw from 2x 8pin pcie. Figure that's going to be possible from an FE on steroids like the SC models, with possibly a 3rd 8pin on the FTW class gpus.