GMM is a general estimation strategy while IV is just an estimator. Many estimators including OLS, GLS, IV, 2SLS... can be seen as special cases of GMM. In contrast to MLE, GMM does not require any distribution assumption but only the existence of certain moments. The efficiency depends on the specification of the weighting matrix. If there is endogenous issue, the IVs are necessary to correct the bias. If the exogenous instruments cannot be found outside, there are some alternative choices. For example, you can decompose the X's into two parts, one of which can be instrument. Or use lag terms of X's. If considering the dynamic panel, you will probably use the Arellano-Bond estimator. That's all I can remember.