All computing devices implement computer arithmetic as a core functionality. Nearly all computing devices moreover implement the same, standardized number formats in order to aid software portability. However, with Moore’s Law ending, we question whether it remains the best approach to achieve high performance and low energy consumption by applying the same standardized number formats for all applications. Among others, the Internet-of-Things (IoT) presents a proliferation of compute devices with a wide variety of functionality, among others in data sensing, data analytics, and inference of machine-learned models. Additionally, security mechanisms, which are essential for internet-connected devices, are compute-intensive workloads. We posit that each of these workloads will see improved performance and energy-efficiency from distinct, specialized computer arithmetic.
The SoftNum Marie Curie Fellowship will explore how to make number formats, generally considered to be hard-wired functionality, software-defined. Software-defined number formats have the advantage of high performance, low energy consumption, and ensure sufficient while not excessive precision. For instance, with SoftNum it will be possible to design a data sensing device that pre-processes data streams at the minimal precision possible to minimize energy consumption, embeds a watermark to prove ownership and origin of the data, all while achieving the required application-level accuracy. This project, if successful, will have a major impact on how we perceive the design of computing devices. It will create an unprecedented degree of virtualization of the core computing hardware, with an estimated system-level speedup of 50% and reduction of energy consumption by 70%. This novel proposition promises to accelerate the Fellow’s career, who brings prior expertise in computer arithmetic to the project, under the mentorship of an expert in transprecise computing and runtime system software.
Fields of science
Call for proposal
See other projects for this call