Typically there are relatively few replicates, so it's difficult to accurately estimate the variance (or dispersion if you prefer that term) on a per-gene level. Consequently, both tools pool information across genes to get more robust measures. Since the reliability of a p-value is determined in large part by how accurate you've estimated variance, this ends up being a major benefit.

Of course if you have a lot of samples (e.g., a thousand) then this isn't really needed.