I'm not sure if there's an exact answer to this question but there are 2 things you need to consider: (1) the number of points in each pooled lag distance, which will be affected by (2) the spacing of lag distances. You need enough points pooled into each (mean) lag distance to estimate the semivariance reliably (some people say >3, preferably a lot more). You need sufficiently small lag spacings (i.e. large enough number of points on your semivariogram) to be able to fit a variogram model reliably (for example, the defaults in R give you about 10-12 points).