Not all rational numbers (within the bounds of a floating point type) can actually be represented in floating-point, which can lead to accuracy problems. Using 64-bit (ie. double) instead of 32-bit mitigates (but does not solve) this problem. In CG we usually don’t care about the extra precision and most programs are written to only use single-precision in most cases. On the other hand, scientific applications or engineering applications might value the extra precision.