Using a simple backend class that wraps any fundamental (built-in) arithmetic type we can measure the overhead involved in wrapping a type inside the number frontend, and the effect that turning on expression templates has. The following table compares the performance between double and a double wrapped inside class number:

As you can see whether or not there is an overhead, and how large it is depends on the actual situation, but the overhead is in any cases small. Expression templates generally add a greater overhead the more complex the expression becomes due to the logic of figuring out how to best unpack and evaluate the expression, but of course this is also the situation where you save more temporaries. For a "trivial" backend like this, saving temporaries has no benefit, but for larger types it becomes a bigger win.

The following table compares arithmetic using either long long or number<arithmetic_backend<long long> > for the voronoi-diagram builder test:

This test involves mainly creating a lot of temporaries and performing a small amount of arithmetic on them, with very little difference in performance between the native and "wrapped" types.