Hmm, I think these results are somewhat suspect. Postgres isn't even really hitting its stride yet at a million rows. You have to start looking at 100GB+ database sizes otherwise most everything fits in working memory, or at least the indexes do.

My experience has been that joins are, in fact, cheap in larger databases but they do scale on the size of tables, so you should be cautious about excessive joining.

In the extreme case, I wouldn't make a table for each attribute unless there were some overriding reason (at which point you've rediscovered column-based databases essentially). I do sometimes go up to 5th or DKNF normal form though, which I think is substantially beyond what most people designing or revamping a database would do.