You don't need double-precision on the graphics card. The way to handle this is to compute camera/view-relative object positions in double-precision and upload those view-relative positions in single-precision to the GPU.
So, if you need a single global origin (i.e. in the middle of the galaxy) and have all other positions (i.e. positions on Earth) relative to that origin then use double-precision on the CPU to compute view-relative positions. This way your camera transformation is not a gigantic translation of the world from origin to the position you are currently at, but will actually only have a rotational component, because you already pre-transformed the translations on the CPU as view-relative positions when rendering the models.