A lot of it will be in the difference in cost. But as a rule of thumb, GW potentials are typically harder, and include more states (semicores).
This need not always be the case, since if you look at the Ag POTCARS, there isn't any difference at all in the recommended ENMIN/ENMAX, partial core radius, or even the states that are considered valence.
In contrast, the Ti POTCARs for plain Ti vs the GW Ti are quite different. The GW POTCAR has a higher cutoff, lower partial core radius, a much higher cutoff requirement for the augmentation charge, etc.
A counter example to the whole 'is probably harder, probably has more states' are the Li/Li_GW or Na/Na_GW potentials, where the GW variant appears to be softer. I do not know why this is. I do not know why this is. There is presumably a reason, but I am not aware of it.
Perhaps a more relevant difference is the pseudo-wavefunction norm, which in this work here, it is derived that you need (at least in the formalism used) an approximately norm-conserving PAW core. This can be seen in the C/C_nc_GW pair, with the nc GW potential having ~3x higher of a recommended cutoff (as well as augmentation cutoff).
From looking at this fairly minimal set of comparisons, the most obvious indicator of the difference in calculation costs from using the _GW potentials will be in the increased cutoff required to converge the core-region charge density/wavefunctions. However, whether those need to be converged when doing a ground state DFT calculation (as opposed to a GW calculation), I do not know.