> It's not like Nvidia completely rebuilt the GPU between revisions

NVidia absolutely changes the assembly language rather dramatically between iterations. That's why the code is in NVidia PTX (or HLSL / Microsoft's DXIL, DirectX Intermediate Language), and then recompiled to actual GPU assembly each time you install a game or run the code on a different system.

I have not seen an instance where a GPU series that has been out for over 12 months has regressed in performance prior to an older one

Perhaps back in the very early accelerator days when everyone was making GPU-specific hacks of Quake 1 but even that was smoothed over by the transition to DirectX/OpenGL/GLiDE at the time