Date of Award
5-2019
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering (Holcomb Dept. of)
Committee Member
Yingjie Lao, Committee Chair
Committee Member
William R Harrell
Committee Member
Jon Calhoun
Abstract
Semiconductor feature size has been shrinking significantly in the past decades. This decreasing trend of feature size leads to faster processing speed as well as lower area and power consumption. Among these attributes, power consumption has emerged as the primary concern in the design of integrated circuits in recent years due to the rapid increasing demand of energy efficient Internet of Things (IoT) devices. As a result, low power design approaches for digital circuits have become of great attractive in the past few years. To this end, approximate computing in hardware design has emerged as a promising design technique. It provides design opportunities to improve timing and energy efficiency by relaxing computing quality. This technique is feasible because of the error-resiliency of many emerging resource-hungry computational applications such as multimedia processing and machine learning. Thus, it is reasonable to utilize this characteristic to trade an acceptable amount of computing quality for energy saving.
In the literature, most prior works on approximate circuit design focus on using manual design strategies to redesign fundamental computational blocks such as adders and multipliers. However, the manual design techniques are not suitable for system level hardware due to much higher design complexity. In order to tackle this challenge, we focus on designing scalable, systematic and general design methodologies that are applicable on any circuits. In this paper, we present two novel approximate circuit design methods based on machine learning techniques. Both methods skip the complicated manual analysis steps and primarily look at the given input-error pattern to generate approximate circuits. Our first work presents a framework for designing compensation block, an essential component in many approximate circuits, based on feature selection. Our second work further extends and optimizes this framework and integrates data-driven consideration into the design. Several case studies on fixed-width multipliers and other approximate circuits are presented to demonstrate the effectiveness of the proposed design methods. The experimental results show that both of the proposed methods are able to automatically and efficiently design low-error approximate circuits.
Recommended Citation
Qiu, Ling, "Designing Approximate Computing Circuits with Scalable and Systematic Data-Driven Techniques" (2019). All Theses. 3114.
https://open.clemson.edu/all_theses/3114