That's correct. You just have an abstraction layer that represents the cube as a specific star schema on a vanilla SQL database.

Other than the way you load data in, what's the difference, though? It's just an implementation detail.

Speed.

You can periodically build a read-only cube that has pre-computed answers to any number that any pivot-table could read from that star schema.

Then, when the multinational supermarket CEO queries for his company's total profit by product line over the past five years, he gets an instant answer.

The other way, the CEO must wait for weeks to get that answer from the RDBMS - or the cube must be replicated the cube inside the RDBMS somehow using tons of materialized views.

For my money, there's no better multidimensional database I've used for operational analytics than SSAS. JasperSoft, Pentaho, Essbase, Cognos, etc all just don't cut it. Essbase if you're doing GL (SSAS doesn't aggregate Parent-Child very well), but SSAS is an incredibly powerful tool. You can use Postgres with it, so you don't need SQL Server RDBMS to go with it, and you can just pay the license for your SSAS server (which needs more memory than CPU, anyway).

But if you've already got SQL Server for the SSAS, would you go through the bother of installing Postgres?

I think that's Microsoft's advantage in the database market: their system has all the tools out-of-the-box.

And the licensing's really not that expensive - especially compared to the heart-attack inducing prices coming out of Oracle, IBM, SAS, Teradata, Kx etc etc etc