Why do OLAP cubes perform better than Relational databases on certain tasks?

My team uses a Multidimensional OLAP database for a lot of it's analytics and predictive modeling. The standard line in all the documentation is "OLAP cubes are better suited for this type of application than relational databases". On the other hand, you frequently hear that a relational database with a star or snowflake schema (sometimes called ROLAP) can do anything that a true OLAP database can do. 1. Why is it exactly that OLAP is better for analytical processing, aggregating, etc... than the relational model? 2. Is it true that ROLAP can do anything that a true OLAP can do, or are there still limitations or performance issues that can only be solved by using a true OLAP db?