Most of the problems with relational databases is that if you have vast amounts of tables that have joins to one or many tables, and if you require to pull off data once off, you will have to optimise your SQL query to make joins efficient.
In NoSQL databases, the main objective was to be able to be fast and scale horizontally. Some do avoid data joins so you would have to do this yourself (by pulling data in memory and do match joins). Facebook's own Cassandra (now an Apache Project) is basically a NoSQL database system which guarantees no single point of failure.
Also, RDBMS indexing is relatively faster (but that can be debatable) compared to NoSQL databases when it comes to indexing large documents.
I haven't played with CouchDB or MongoDB so I can't compare them. All I know is some do joins in memory (like Redis) which in effect means, pulling all data from database to memory (RAM) and doing joins.
I don't know if that's what you're looking for.