toldfe for a primitive cell vs a supercell?

[fhssn1]

Do we relax the toldfe value (for SCF run) when we use a supercell compared to when we use a primitive cell?

E.g., if we are using toldfe of 1e-12 for a 4 atom primitive cell, and now we need to do a 2x2x2 supercell calculation (32 atoms). The system is 8 times bigger, so we can use 8e-12 instead and we should get equivalent convergence?

Secondly, I'm a little uncomfortable with using values in the range 1e-11, 1e-12. I know a comment in tutorial files says '1e-6 is too large for serious/production runs'. So I use 1e-9 or 1e-10 instead. And that means the ETOT progression shows 9 and 10 decimel places matching before the SCF run stops. That "looks" like a very high level of convergence. Do we really need to go 1e-11, 1e-12 instead? what about 1e-13, 1e-15? (As a numerical person, I know we can't go too low because then we would start having round off errors. In addition, such low values assume implementation of very robust numerical algorithms. So I'm wondering if that is kept in mind when such recommendations are made. Just trying to learn).

[ebousquet]

Supercell or not, the convergence criteria are the same.
Regarding the best value to use, I recommand you to test with the physical properties you want to calculate, i.e. test how your property vary with told-something flag, maybe you don't need a very high fine tuning of the SCF energy. It is system, pseudopotential and property dependent, so you have to check according to waht you want.
Cheers,
Eric