Most GPUs use the IEEE 754 single-precision floating-point format, known as FP32. The thing about using fixed representation for floating point values is that sacrifices have to be made. As you get further from 0.0, floating point gets less and less precise until…well these things happen.

Here’s what happens when you start translating objects far away from (0.0, 0.0, 0.0). After about 1km, things start getting imprecise enough to notice.

Here’s a good explanation.