Force the use of the fast multipole method [Greengard87, Greengard88, Greengard94, Burant96, Burant96a, Burant96b, Strain96, Millam97, Izmaylov06] if possible. The use of FMM is automated in Gaussian 16. Gaussian 16 generally turns on the FMM facility when using it provides even a modest performance gain (say, 1.2x). FMM is enabled for nonsymmetric molecules with 60 atoms or more for both Hartree-Fock and DFT. For molecules with high symmetry, FMM is enabled for Hartree-Fock and hybrid DFT above 240 atoms and for pure DFT above 360 atoms. For molecules with low (but non-zero) symmetry, intermediate thresholds are used. You will begin to see substantial performance improvements (2x or better) for systems that are twice as large.
Of course, the exact results will vary from case to case (compact systems show the least speedup; stretched out linear ones the most), but the defaults are very unlikely to enable FMM when it has a negative effect on performance and are also as unlikely to fail to enable it when it would be worth a factor of 1.5x or more. Thus, users are unlikely to need to control FMM by hand except for some very unusual special cases, such as nearly linear polypeptides and long carbon nanotubes.
The options to FMM are described in Program Development-Related Keywords.
Energies, gradients and frequencies for HF, pure and hybrid DFT. This keyword may also be used within method specifications for ONIOM layers.