Shadow maps are indeed one of the rendering techniques where hardware can become more easily fillrate bound. It's particularly a case of depth-only rendering where shadow texel projection is tested against the depth buffer from the persepctive of light sources. A probable reason why it is optimal to combine UE5's Nanite meshes with higher resolution virtual shadow maps is so VSM can reuse the renderer to potentially software rasterize the shadow maps thus bypassing the limits of fixed function hardware depth testing!
For the G-buffer depending on the number of render passes, we could see fillrate cost adding up per pass as games can frequently modify the G-buffer dozens of times for a number of reasons (UI/screen-space graphical effects/decals/post-processing/etc). The alternative (if possible) to drawing (draw calls) these render passes is to use compute dispatches (no draw calls) to render these pass instead. No draw calls = No ROPS!
OCing takes it to 5080 levels, but you can also overclock the 5080, that's true 
