Most games I know of generally do this, I infer:
use 1 or 2 shadowmaps, which are grid-snapped while travelling with the camera (2nd is to gradually lerp between two, for smoother transition).
1 reflection-tex, unless stenciling
Only a handful cubemaps, if more than one. One dynamic cubemap, max (in racing sims mostly).
glGenerateMipmap seems slow, ime. And the blocky results are less than desirable to a manually gauss-blurred downsampling ping-pong.
I notice no performance drops on RG11B10f formats vs RGBA8, in msaa/csaa.
/no experience in shipping a hw-accelerated product, take with a handful of salt/