Ultrasoft pseudopotential (USPP), in general, requires a lower energy-cutoff to the plane waves when compared to Norm-Conserving ones (NC). On the other hand (at least with Quantum ESPRESSO) the charge-density cutoff should be bigger for the USPP, where it is always kept 4 times the energy cutoff for the NC. Apart from that, which other characteristics should be taken into account when choosing the pseudopotential? Are there specific properties more or less sensitive to one or another?
Three criteria to consider are
- Performance and size of basis set USPPs generally require a lower plane-wave cut-off and smaller basis, but a larger density grid. However performance may not be straightforwardly related to cut off as there are additional terms to compute which may have a significant computational cost.
- Accuracy USPPs usually have 2 (or more) projectors per angular momentum state, which linearizes the energy. Traditional Troullier-Martins and Rappe/Cambridge NCPs had only one and were consequently less accurate. Do beware of older generations of NCPs for this reason, but recent multi-projector NCPs, nowadays known as "Optimized Norm-Consetving Vanderbilt" are nearly as accurate as USPs. See the Delta project website for some examples.
- Implementation restrictions and constraints Some functionality may be implemented only for NCPs, so USPs may not be used.
The short answer is DEPENDS on your system and property of interest you are after (among other things).
Anyway, I see the order of physical basis/complexity increasing as NC --> USPP --> PAW. But I may be wrong here.
However, you should look at https://www.vasp.at/vasp-workshop/pseudopp2.pdf and the original USPP paper by Vanderbilt https://journals.aps.org/prb/abstract/10.1103/PhysRevB.41.7892