I'd be curious to know in what contexts anyone would not want transactions.
I readily see the case for memcached or equivalent when it comes to caching -- it is very valid, and indeed useful, when slightly outdated data is perfectly acceptable. I can even picture how you might be using MongoDB to do the same, even if you admittedly have me wondering how you're invalidating the mess.
For a persistent data store, however, I'm honestly at a loss. In my own experience, transactions are needed as soon as there's a remote possibility that a concurrent write may occur. Even embedded systems need them, when you're threading statements concurrently for performance reasons, or when you're subsequently merging local data with another node. (See CoreData/iCloud bugs for what occurs when you ignore ACID in the latter case.)
In your own question you kind of hint at it. You need a tuned database. NoSQL lowers the barrier to entry for fast persistent data storage with replication. NoSQL doesn't replace SQL databases, it's looking to optimise a use case where transactions are not required.
Personally I use Mongo a bit like a cache, sitting in front of a SQL DB. Things that need to be ACID are handed back and forth. Those that don't are kept in Mongo and I get the best of both worlds.