It ensures the variance is independent of the mean.

It forces the parameter to be normally distributed.

It results in the sufficient statistic of the distribution being equal to the linear predictor $X\beta$.

It eliminates the need for maximum likelihood estimation.