The difference is approximately normal if $n_1, n_2$ are large.

The variance of the difference is $\frac{\sigma_1^2}{n_1} + \frac{\sigma_2^2}{n_2}$.

It requires the underlying populations to be normal.

It only applies if $\sigma_1 = \sigma_2$.