Depends on the application really. The operation itself would surely be slower than a specialised mmul, but there can be big advantages to keeping it inside your database if you can phrase most of your problems in terms of simple linear algebra operations.
In other words, numpy is damn-fine with matrices, but if you can replace DBMS->network->ORM->numpy->ORM->network->DBMS with just some SQL, it's pretty clear which will be faster.
On the other hand, if your data isn't already in an RDBMS in matrix-like tables, there is little point in moving it to one just so you can use tricks like the one mentioned in this article.
It's not a speed optimization. It just allows you to process a much larger dataset than you can fit in main memory. It will be slow because it's hard disk I/O, but you can do a sparse matrix multiplication on terabytes worth of data in a database with SQL. You couldn't do that in Python or Java without running out of memory, unless you used a MapReduce job distributed to many worker machines--and then you'd get a bill for the EC2 time or whatever.
In other words, numpy is damn-fine with matrices, but if you can replace DBMS->network->ORM->numpy->ORM->network->DBMS with just some SQL, it's pretty clear which will be faster.
On the other hand, if your data isn't already in an RDBMS in matrix-like tables, there is little point in moving it to one just so you can use tricks like the one mentioned in this article.