Sherief Reda - Approximate Circuits - Methodologies and CAD

This Springer imprint is published by the registered company Springer Nature Switzerland AG

The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Approximate computing has emerged as a new paradigm to reduce the resources (e.g., design area and power) required to realize digital systems at the expense of a negligible or small amount of reduction in quality-of-results or accuracy. This trade-off between resources and accuracy is specially relevant for a large class of data-rich applications such as machine learning and multimedia processing that offer inherent error resiliency. This chapter overviews the main technical themes in approximate circuit design methodologies. We elucidate various application domains that are most suitable for approximate circuits, and we then describe a number of error metrics that can capture the quality of results of the applications. We overview the main four technical themes of this book, which are (1) design of approximate arithmetic building blocks, such as adders, multipliers, and dividers, (2) circuit synthesis techniques for arbitrary logic circuits, (3) approximate accelerator design for a number of applications, including deep learning and video coding, and (4) approximate circuit techniques for general-purpose central-processing units and graphical processing units.

With the emergence of more and more complex applications in domains like machine learning and multimedia processing, the overall computational workloads of the applications and their respective energy consumption are on the rise. Applications from the domains of Internet-of-Things (IoT) and cyber-physical systems (CPS) that involve huge amount of data analytics require a significant amount of computational and thereby energy and power resources. To overcome these escalating challenges, technology scaling has played a vital role in the past few decades; especially the accelerator-based computing has shown promising results for high energy efficiency. However, with the diminishing returns of technology scaling, alternative computing paradigms have to be considered for alleviating the resource requirements of the applications and for providing near-optimal level of performance and energy efficiency.

Approximate computing (AC) is one such paradigm that leverages error resilience characteristics of applications for improving the overall resource efficiency of the systems. Traditional techniques, like power gating, dynamic voltage and frequency scaling (DVFS), and power-/energy-aware application mapping, which are popular and have been widely used for achieving significant efficiency gains, are not sufficient enough to meet the growing computing-efficiency demands and can only offer improvements to a limited extent. However, relaxing the bounds of precise computing can open new horizons for the designers by offering (design- and run-time) trade-offs that were never possible using conventional techniques. Figure shows a comparison between traditional and emerging approximate computing-based HW/SW computing stacks.

Fig. 2

Comparison of hardware and software stacks in traditional and approximate computing paradigms. The highlighted blocks in the stack illustrate the focus of the book (Adapted from [13].)

This book is focused towards introducing the most prominent techniques proposed for employing approximations at the hardware level for achieving high gains in terms of area, power, energy, and performance efficiency. The book is composed of several parts that cover approximate circuits as well as the methodologies for designing approximate hardware. More details related to parts are available in section of this chapter.

Application Domains

Approximate computing techniques benefit circuits and systems from application domains that have inherent error resilience as described in section . We describe in this section specific examples of these application domains.

1. Big Data Applications. Big data applications have inherent noise in their data collection process, where large volumes of data often include statistical anomalies. Furthermore, the collected data can be based on inputs from users which can be subjective and can include erroneous or unexplainable data points. Furthermore, big data processing often relies on machine learning techniques, where the results are not 100% accurate anyway.