To continue on my previous post,
Second, point reasoning. You ideally want your multidimensional structures to be dense as possible. Every combination of every member of every attribute hierarchy to be referenced somewhere in your partition stores. That's not going to happen obviously as for a retail system, not every customer buys every product on every day (and in Steve's Case, every second). But to combine date and time just increases the sparcity of the multidimensional database along a dimension that is most likely going to be used to partition a database of any size whatsoever. It would just be a huge mistake to do so. Like said earlier, separate date and time dimensions for 10 years worth of data at the minute resolution results in a dimension with over 5 million (where'd I get 13M from?) stored members while separate dimensions results in a date dimension with 3650 members and a time dimension with 1440 members. I don't need benchmarks to tell me the separate dimension approach will best the consolidated dimension approach in every aspect; dimension processing, aggregation processing, ease of maintenance. Everything.
Thirdly, say you have a referenced date in your fact table but no foreign key to your date and time dimensions. You could use Steve's approach to reference a virtual TIME (and DATE) dimension using recursive CTEs to force into SSAS. But wait. If I convert a datetime type to an int [CAST(FactDate AS int)], excluding the time portion (assumes its AM), gives me the number of days elapsed since 1/1/1900. Hey, why not use that as the surrogate key of your date dimension! (Which is exactly what I and others do) I don't even have to do a look up for a date key at all in my ETL and if I haven't already, I could create a calculated column in my DSV. Similarly, I could use the number of minutes (or seconds) that have elapsed since midnight as the surrogate key of my time dimension so that I don't have to do a lookup for the Time Key either. (And could still create a calculated column for that key in my DSV) And because of star joins and the small size of the date and time dimensions, analysis of events between arbitrary date/time periods is lightning fast. (And not nearly as "evil" as the CTE Steve purposes as the replacement of the time dimension)
So to summarize, I still would give this article zero stars. I think it subtracts, not adds, to the body of knowledge.