That's pretty hard to give a straight number on, for one thing, you'd only know after the run, if your baseline was good enough, and it really does depend on your application. FWIW, in some applications I've worked with, an impurity peak can be ignored if it's less than 5 times the noise, for other applications, in the same group, all peaks were significant, provided they are greater than the noise over a respectable interval. But that only helps you in a roundabout way -- noise and drift must be significantly less than what you expect to be analyzing, but you don't know what that is, until you try.