Perhaps the biggest issue currently is that Nvidia in recent generations of their video cards have decided double-precision is a “pro” feature and consumer cards have no hardware acceleration for DP so things end up 24-32 times slower when you try to do DP calculations in CUDA compared to SP, so you would lose Cycles GPU compute on consumer Nvidia cards if Cycles switched to 64-bit double-precision floats. If AMD starts offering cheap DP performance then maybe Nvidia will have to change its ways.
Single precision basically limits your 3d universe to 33,554,433 “places” in each dimension, so take the maximum dimension based on your unit scale and divide by 32M and that gives you some idea of the smallest distance that can exist between two separate points. When that becomes significant then the math starts to go pear-shaped, especially in rendering.
So in a cubical volume 2km on a side (+/- 1km from the origin), you get something like 16 places per mm that you can put a vertex, so you can do pretty finely detailed modeling and never notice. If you want to work on a sphere the size of Earth, now you only have places to put vertices about every 1/3 of a meter though space. If you’re modeling large structures to be seen from a distance then maybe that’s not horrible, but try modeling a 40cm teapot sitting on the ground and you’ll have a bad day.
And things like procedural textures become a pain to work with as you end up having to multiply everything by the scene’s unit scale, etc.
As usual all of the above is subject to being nonsense if I messed up the math which is entirely possible 