reference request - Numerical algorithms on mixed-precision computational models.

I wonder whether this approach is worthwhile on modern hardware. There's not as much difference between integer, float, and double operations as there used to be.

One way it may matter, however, is saving space. With smaller data types, you can fit more numbers at a time into cache, and that may give you a speed up. The time savings doesn't come from arithmetic operations but from faster access to the data if you can avoid a cache miss.