Software-Defined Number Formats: Bridging the Gap between Accuracy, Performance and Security

Nearly all computing devices implement the same standardized number formats to aid software portability. However, with Moore’s Law ending, we question whether it remains the best approach to achieve high performance by applying the same standardized number formats for all applications. In particular, the efficiency promises of low-precision compact number formats imply that a single format is no longer sufficient for all applications, and single application may even benefit from using multiple distinct compact formats. As a result, relying on a single (or limited) hardwired format is no longer effective. This poster presents our exploration of software-defined number formats, a novel approach to make number formats, generally considered to be hard-wired functionality, software-defined. These formats are designed specifically to meet the numeric needs of the computation and are more compact representations of the data. They reduce memory footprint and memory bandwidth requirements without sacrificing accuracy which results in a speedup compared to conventional floating-point formats.