Approximate Computing for Low-Power Many-Core Processors

低功耗众核处理器的近似计算

• 批准号: RGPIN-2018-03854
• 负责人: Baniasadi, Amirali
• 金额: $ 1.68万
• 依托单位: University of Victoria
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=752031
• 关键词: Approximate; Computing; Low; Power; Many

Developing more efficient computing systems is a critical goal of the high-tech industry. To this end, designers have suggested hardware (i.e., energy efficient components and systems) and software methods (compiler solutions for generating energy efficient code). Developing efficient computing systems not only has resulted in environmental benefits, but has also helped reducing production cost and complexity and achieving longer battery lives. Recent studies however, have shown that the same level of efficiency is not sustainable and that further achievements in developing low-power solutions face difficulties imposed not only by technology limitations but also application requirements.Traditionally, applications running on computing systems have required intact accuracy. In recent years many studies have analyzed computing at the application level and have shown that many applications can tolerate an acceptable level of inaccuracy and approximation in their computations.Approximate computing builds on this observation and is a promising solution to maintaining energy efficiency in the future. Conventional low-power design explored a two dimensional space where trade-offs between power and performance were explored with the goal of maximising energy savings while maintaining performance. With the inclusion of approximation, this design space has been extended to a three dimensional one where accuracy can also be traded for higher efficiency. This extension motivates us to pursue new opportunities and further optimizations. This research will deliver new ways to explore and build approximate-aware systems relying on a less aggressive approach to computing but still capable of delivering acceptable results.We propose a deep analysis of General Purpose Graphic Processing Units (GPGPUs) anddeveloping both hardware and software solutions to reduce energy consumption and improve efficiency while maintaining accuracy within acceptable limits. While our hardware solutions will focus on designing new and efficient Graphic Processing Units, our software solutions will use a compiler based approach to identify opportunities to achieve low-complexity computing. Our past experience with hardware and software optimizations, application behaviour analysis and our familiarity with the tools available provide us with the required skills to develop low-power approximate-aware systems.Our proposed research makes important contributions to the Canadian society. First, the proposed program aims at developing "greener" computing infrastructures where power hungry GPGPUs are replaced with low-power alternatives. Second, the resulting knowledge belongs to an area critical to Canada's future leadership in advanced technologies. Many Canadian companies can benefit from the findings of the proposed research and the resulting HQP training.

开发更高效的计算系统是高科技行业的关键目标。为此，设计者建议硬件（即，节能组件和系统）和软件方法（用于生成节能代码的编译器解决方案）。开发高效的计算系统不仅带来了环境效益，还有助于降低生产成本和复杂性，并延长电池寿命。然而，最近的研究表明，相同的效率水平是不可持续的，并且在开发低功耗解决方案方面的进一步成就不仅面临技术限制，而且还面临应用要求的困难。近年来，许多研究分析了应用级的计算，并表明许多应用程序可以容忍其计算中的可接受水平的不准确性和近似性。近似计算建立在这一观察的基础上，是一个有前途的解决方案，以保持能源效率在未来。传统的低功耗设计探索了一个二维空间，其中探索了功率和性能之间的权衡，目标是在保持性能的同时最大限度地节省能源。随着近似的加入，这种设计空间已经扩展到三维空间，在三维空间中，精度也可以换取更高的效率。这种扩展激励我们追求新的机会和进一步优化。这项研究将提供新的方法来探索和建立近似感知系统依赖于一个不那么激进的方法来计算，但仍然能够提供可接受的结果。我们提出了一个深入的分析通用图形处理单元（GPGPU）和开发硬件和软件解决方案，以减少能源消耗，提高效率，同时保持在可接受的范围内的准确性。虽然我们的硬件解决方案将专注于设计新的高效图形处理单元，但我们的软件解决方案将使用基于编译器的方法来识别实现低复杂度计算的机会。我们过去在硬件和软件优化、应用程序行为分析方面的经验以及我们对现有工具的熟悉程度为我们提供了开发低功耗近似感知系统所需的技能。我们提出的研究为加拿大社会做出了重要贡献。首先，该计划旨在开发“更环保”的计算基础设施，用低功耗的替代品取代耗电的GPGPU。其次，由此产生的知识属于加拿大未来在先进技术方面的领导地位的关键领域。许多加拿大公司可以从拟议的研究结果和由此产生的HQP培训中受益。