Right-Sizing Precision: Save Energy, Power, and Storage

Most computer programs use far more precision than needed, because programmers find it easier to use excess precision as insurance against rounding or overflow errors instead of analyzing what precision is actually needed. Sometimes we might actually need more precision. We are attempting ever more ambitious calculations with ever less attention to numerical analysis. While this may sound like an arcane issue, it hits the energy and power budget hard because so much energy is wasted moving excess bits between memory and processor. There are a number of interesting approaches for "right-sizing" the precision we use that don't require programmers to become experts at numerical analysis. A particularly intriguing approach is that the way we represent numbers in computers could be made self-descriptive and accuracy-aware, bringing us closer to the goal of automatic and economical control of rounding errors.

Speaker: John Gustafson, AMD