[corysama]

Lots of good answers in the comments here.

Similar to the PathEngine concept, when I compress 3D meshes to 6 bytes per xyz, I could use float16. But, there's no point to having most of the precision focused in the center of the mesh. The verts are usually fairly evenly distributed if not biased towards the outside edge of the bounding box.

Bit-level, cross-platform floating point determinism is harder than it should be. That and associativity is important in physics, financials, replays and distributed simulation.

The PlayStation1 and GameBoyAdvance didn't have FPU coprocessors. Software floating point was completely impractical. So, you pretty much had to use fixed point. To the point that the PS1 did have a fixed-point vector math coprocessor! :D