The three inputs “CUTOFF”, “NGRIDS”, and “REL_CUTOFF” are important for accuracy and efficiency. The problem is that both, “accuracy and efficiency” depend on the combination of chosen values. See also the FAQ about cutoffs.

In an optimal setting you can proceed the following way:

1) choose a high enough value for REL_CUTOFF

 this is system and basis set independent. It determines the
 minimal cutoff used to bring a Gaussian function on the real space grid.
 High values mean high accuracy, low values faster execution.
 

2) choose your cutoff and number of grids. This will determine the real space grids

 for your calulcation, e.g. cutoff=400 ngrids=4
 grid1 = 400, grid2 = 200, grid3 =100, grid4 = 50 (not actual numbers)
 The cutoff should be high enough to take on the Gaussian with the largest exponent,
 i.e. exponent*REL_CUTOFF <= cutoff (where exponent is for a product Gaussian)
 If you choose more grids the mapping (Gaussian->Grid) is better but you have more
 overhead from more grids (mostly FFTs).