hmmm. there is no 'hard and fast' rule. only general guidelines. and of course it depends on what you are using the information for.
there are two standard deviations that are calculated form the parts.
the first is the measurement repeatability. there are traditionally either 2 or 3repeated readings per part/operator combination, so n=2 or n=3. The number of parts that are measured effects the precison of the estimate of the repeatability. in general if you have 'several' parts that span the range of part to part variation for this estimate you are probably OK. I've been successful using 6 parts (3 at the low end and 3 at the high end).
Then there is the estimate of total variation that is used ot estiamte part to part variation. if you have historical information on the total variation, fewer parts are probably sufficient as all you really need is the measruement error (repeatability) variation. but if you need to use the parts from teh study for this calcualtion you can get really misleading results if you use smaller numbers of parts. The statistical problem is that you are estiamting the standard deviation of the total variation using only a few parts. they can not be 'random' or representative of the actual distribution and so you will most likely OVER estimate the total variation if you take samples that span the range, or UNDER estimate it if you do take a random sample becuase you will likly gets parts that are all close in value. estimates of standard deviations typically require larger sample sizes to have any reliable precision.
for general studies I use 30 parts and 2 repeats.