We have done tests (http://www.anchormodeling.com/wp-content/uploads/2010/09/Performance-Testing.zip) that show better performance in an AM model with a large number of attributes than in a 3NF model, when you don’t fetch all attributes at the same time (select * from …).
If you do select * you will not use the benefit of table elimination and your performance will most likely be worse than a 3NF model, but it could still be better if there are many versions of your data and attributes are historized asynchronously with respect to each other, or if you have where-conditions that are very selective and over columns for which the database has collected statistics, or when data is sparse.
AM models are very good when you have sparse data because you will only store data that has been specified. Null values correspond to the absence of rows in Anchor Modeling, resulting in less data to scan during query execution.
If your data is dense, static, and you often need to fetch all attributes at the same time in queries that have few or no conditions (and therefore result in big joins), then performance will suffer.
