The documentation states that you do get unlimited queries, but you're still limited by query rows. As such, you should attempt to optimize your code to use as few queries as possible.
For example, here's the output from one of our metadata tables that has 1,224 rows, when we query it once:
21:20:52.1 (241740102)|USER_DEBUG||DEBUG|Limits.getQueryRows: 1224
21:20:52.1 (241758076)|USER_DEBUG||DEBUG|Limits.getQueries: 0
And here's the output from doing 40 queries:
21:23:18.1 (468898239)|USER_DEBUG||DEBUG|Limits.getQueryRows: 48960
21:23:18.1 (468917235)|USER_DEBUG||DEBUG|Limits.getQueries: 0
And 41 queries:
21:23:47.1 (527992170)|EXCEPTION_THROWN||System.LimitException: Too many query rows: 50001
21:23:47.1 (528427073)|FATAL_ERROR|System.LimitException: Too many query rows: 50001
So, while you do get an "unlimited" number of queries to custom metadata, keep in mind that you are still using CPU time, heap limits, and row limits. Query only what you need to, and cache your results for performance reasons.
In practice, this limit means you couldn't actually get away with querying custom metadata more than 50,000 times at one row a piece.
Also, based on testing, there's some CPU time used for each query, so the net result is that you'll eventually hit CPU time limits. In my test, hitting my 1,224 entry table in a loop resulted in a CPU time usage of about 5.5 ms per query. This may seem like a small value, but multiply that by 1,000 and you're already at 5.5 seconds of CPU time just to process those records.