Depends on what you want to accomplish really. There would be significant difference between several runs so calibrating one chip and using the same config for all the chips coming from the same run would eliminate differences between several runs. But if the temperature sensor you're using is sensitive to process variations or mismatch in a single run will still be there. To eliminate everything you need to calibrate all of them.
Also it heavily depends on requirements, like how much accuracy is needed, how susceptible is the sensor to process variations. Even with these taken into account real world results may differ, so you need to make some tests to see how much variation you get with each method (not being able to calibrate properly is also another deal).
There's no way anyone can answer this question for you as it depends on your requirements. I can only point out that calibrating all of them will yield better results which is obvious (if you're good at calibrating of course).