Approximate and Stochastic Computing Systems

当前的计算机系统仍然消耗大量的电力，无论系统的大小：给它一个智能手机，一台个人电脑或一台计算机服务器。原因有两方面：1)基本电子器件的尺寸很小（在纳米尺度上），容易受到制造和环境因素的影响而产生变化和暂时的误差，因此需要比必要的电压和电流更大的电压和电流来确保运行的可靠性；2)计算应用越来越复杂，例如多媒体，它涉及大量的算术运算。因此，能源效率已成为当前计算机系统的首要关注点。可靠性和能源效率之间的冲突似乎是不可避免的，并提出了重大的设计挑战。然而，另一方面，许多计算任务表现出不精确容错或容错的共同特征，特别是在中间计算过程中。在许多新兴应用中，如图像/视频处理、模式识别和机器学习，它是一个特别重要的功能。本研究计划的目标是开发一种新型的计算系统，该系统采用近似计算（AC）和随机计算（SC）技术来实现节能和高性能的处理。AC利用许多应用程序中的错误恢复能力，并采用深思熟虑和确定性的设计来提供不精确但足够好的结果，而SC使用简单的硬件和随机二进制位流来在计算结果中产生有意义的统计数据。对于交流的基本电路元件的开发以及SC的基本构建模块的设计和实现已经投入了相当大的努力。然而，将各种电路元件集成到交流和/或SC系统中以实现低功耗和高性能的操作是一个挑战。为了应对这一挑战，将各种交流和SC电路元件有效地集成到一个更大的系统中将是本研究的主要重点。该设计将针对嵌入式和移动系统中的传统数字信号处理，以及新兴的以大脑为灵感的计算系统，这些系统探索神经元的组织和功能，并在不同层次和抽象层次上连接突触。将考虑两类基本应用：1)多媒体，包括图像、音频和视频处理；2)使用有效的机器学习模型（包括神经网络）进行图像和语音识别。解决一个根本具有挑战性的问题，即计算机系统的能源效率，这个研究项目将为电子和信息产业产生有用的结果。它还将为具有加拿大经济长期增长所需要和至关重要的技能的高素质人员提供宝贵的培训机会。

Current computer systems still consume a significant amount of power, regardless of the size of the system: given it a smart phone, a personal computer or a computer server. The reason is two-fold: 1) the tiny sizes of basic electronic devices (at a dimension of the nanometer scale) make them susceptible to variation and temporary errors due to manufacturing and environmental factors, so larger-than-necessary voltages and currents are required to ensure the operational reliability; and 2) computing applications have increasingly become complex, such as multimedia, that involve a significant number of arithmetic operations. As a result, energy efficiency has become a paramount concern for current computer systems. The conflict between reliability and energy efficiency seems to be inevitable and presents significant design challenges. On the other hand, however, many computing tasks show a common characteristic of being imprecision-tolerant or error-resilient, particularly during the intermediate computing process. It is an especially important feature in many emerging applications such as image/video processing, pattern recognition and machine learning. The objective of this research program is to develop a new class of computing systems that employs approximate computing (AC) and stochastic computing (SC) techniques for energy-efficient and high-performance processing. AC leverages the error resilience in many applications and employs deliberate and deterministic designs to deliver imprecise but good-enough results, whereas SC uses simplistic hardware with random binary bit streams for producing meaningful statistics in the computed result. Considerable effort has been devoted to the development of basic circuit elements for AC and to the design and implementation of basic building blocks for SC. A challenge, however, is to integrate various circuit components into an AC and/or SC system for low-power and high-performance operation. To address this challenge, the effective integration of various AC and SC circuit components into a larger system will be the primary focus of this research. The design will be aimed at conventional digital signal processing in embedded and mobile systems, as well as emerging brain-inspired computing systems that explore the organization and functions of neurons and connecting synapses at different levels of hierarchy and abstraction. Two essential classes of applications will be considered: 1) multimedia, including image, audio and video processing; and 2) image and voice recognition using effective machine-learning models, including neural networks. Addressing a fundamentally challenging issue, i.e., energy efficiency in computer systems, this research program will produce useful results for the electronics and information industry. It will also provide valuable training opportunities for highly qualified personnel with skills demanded by and crucial to the long-term growth of the Canadian economy.