Half Precision

I was looking up floating-point numbers yesterday just out of curiousity and I found out that the IEEE754 standard for floating-point numbers includes half-precision numbers. These mean exactly what the name says – it provides half the precision of single-precision floating-point numbers. I had once read an article which argued that for most normal applications, we do not require high precision numbers. An example application would be voice applications, that only require a quarter-precision value to work.

A half-precision number is a 16-bit representation for a floating-point. It has the ability of providing a larger range of values than a 16-bit integer representation but is less computationally expensive to work with, than a 32-bit single-precision representation and uses up half the memory space. According to the Wikipedia article, it was mooted by the graphics industry for working on graphics applications. This proves that lower-precision numbers can actually be ‘good enough’ for certain applications.

I just thought that I should note it down here. It will come in useful at some point in the future, when I work on enhancements to my processor.