They've got an elegant and, as far as I know, unique modeling approach that avoids the first-touch/last-touch/path-analysis pitfalls.
Also, would I need an ad server in place to properly attribute display or can they handle that on their own?
It's a unique approach in that they don't need a lot of granular attribution data, but instead use higher level aggregate data. As a result it's easier to integrate with all kinds of media campaigns.
You need to be able to provide impressions over time for your early funnel ads and conversions over time for your late funnel touchpoints. They use some econometric time series techniques to analyze and estimate the impact so you can focus on whats working.
GA's attribution tools are fun to play with, but at the end of the day they are still static models that don't evolve over time as they collect data.