CAREER: Approximate Computing Systems for Future Teraflops Workloads

在过去的四十年里，计算机的性能有了巨大的提高，这使得无数的应用程序在我们的日常生活中发挥了重要作用。然而，如果在提高计算功率效率方面没有显著的创新，仅靠持续的半导体器件缩放将无法为未来的计算能力提供足够的性能。针对新出现的挑战，本研究提出了革命性的计算范式，以极其有效的方式处理、理解和使用大量数据。该方法的新颖之处在于，通过利用通常使用物理现象的复杂概率数学模型的未来工作负载的特征，开发具有软件支持的高性能计算系统。在这种应用中，近似计算通常可以得到令人满意的结果。同时，它可以大幅降低功耗或提高性能，用简化电路代替完整的逻辑功能，模拟粗略计算的功能。为了将近似计算概念扩展到更通用的计算系统，提出了以下整体方法：1)智能微架构，在编译器支持下识别代码的正确性-非关键区域；2)近似计算引擎以有效地执行这些区域的代码功率和性能；3)高级形态原语，以更高的功率和性能效率处理大部分工作负载；4)灵活的体系结构允许程序员和用户以效率换取计算质量。拟议的研究将对计算机体系结构、电路和编译器社区产生具体和重大的影响，因为它需要分析有趣的和有代表性的工作负载；实现最先进的电路、架构和编译基础设施；并发明了强大而有用的评估方法。由于大部分开发和研究工作将由研究生进行，工业界和学术界都将受益于受过良好教育和训练的雇员，以及当学生毕业并开始在其他地方就业时的直接技术转让。最后，这项研究的成功将极大地有利于我们提高人类的能力。我们在科学、技术、商业、医学以及几乎所有人类努力的领域的集体知识，允许计算性能的显著提高。

The performance of computers has improved tremendously in the past four decades, which has enabled innumerable applications that have major roles in our daily lives. However, without dramatic innovations in improving power efficiency of computing, the continued semiconductor device scaling alone will fail to provide sufficient performance for the future computing capabilities. For the emerging challenge, the proposed research investigates revolutionary computing paradigms to process, comprehend, and use abundant data in an extremely efficient way. The novelty of the approach lies in developing power- and performance-efficient computing systems with software support by exploiting the characteristics of future workloads that often use complex probabilistic mathematical models of physical phenomena. In such applications, approximate computing can often result in satisfactory outcomes. Meanwhile, it can dramatically decrease power consumption or increase performance by replacing complete logic functions with simplified circuits that mimic the functions for rough calculations. To extend the approximate computing concept to more general-purpose computing systems, the following holistic approaches are proposed: 1) intelligent microarchitectures to identify correctness-non-critical regions of code with compiler support; 2) approximate computing engines to execute such regions of code power and performance efficiently; 3) high-level morphic primitives to process a large fraction of workloads with orders-of-magnitude greater power and performance efficiency; and 4) flexible architectures to allow programmers and users to trade the quality of computing with the efficiency. The proposed research will have a specific and significant impact on the computer architecture, circuit, and compiler communities since it requires analysis of interesting and representative workloads; realization of state-of-the-art circuit, architecture, and compiler infrastructure; and invention of powerful and useful evaluation methodologies. Since most of the development and research work will be conducted by graduate students, both industry and academia will benefit from well-educated and trained employees as well as direct technology transfer when students graduate and begin employment elsewhere. Finally, the success of this research will tremendously benefit our ability to advance human?s collective knowledge in science, technology, business, medicine, and virtually every other field of human endeavor by allowing remarkable improvement in computing performance.