SHF: Small: Collaborative Research: Integrated Framework for System-Level Approximate Computing

Northeastern University, Boston MA

Investigators

Abstract

Nanocomputing is encountering fundamental challenges with respect to performance and power consumption; it requires different computational paradigms that exploit specific features in the targeted set of applications as well as an integrated framework for assessing the interactions between hardware and the processing algorithms (software). Approximate (inexact) computing has been advocated as a novel approach for nanocomputing design. Approximate computing generates results that are good enough rather than always fully accurate and correct outputs. Recent advances at circuit level have shown that there is an urgent need to investigate and enable at system-level the flexible utilization, improvement and close monitoring of approximate resources; this allows the efficient and integrated interaction of algorithms and hardware to meet the multiple and often conflicting figures of merit of high performance, lower power consumption and reduced inaccuracy. The goal of this project is to develop approximate computing systems that are capable of adjusting performance by exploiting relationships between hardware and software (referred to as intra-level) in different applications such as cognitive processing, DSP, big data and scientific processing for which data can be adaptively utilized and manipulated. This project is an organized effort that combines recent advances in technology with architectural enhancements into an integrated framework for approximate computing that will tackle the critical challenges of emerging computer designs in a comprehensive manner. This framework consists of new functional and computational primitives of hardware resources and related algorithms to allow an evaluation at system-level to meet the desired metrics for approximate computing. Intra-layer relationships such as number representation (such as floating point and logarithm) and accuracy by employing dynamic approximation schemes and data remediation for both communication and computing are also analyzed. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.