SHF: Large: High-Performance, Low-Power, Self-Evolving Integrated Systems through Statistical Learning in Chip (SLIC)

尽管尽了最大努力，但电子芯片的制造具有独特的个性，这源于无法精确制造其底层电路，也无法创建用于控制由此产生的不确定性的软件。使用复杂的测试方法来确定性能最佳的系统是可能的，但这将导致巨大的成本。电子芯片的个性进一步受到其环境和用途的影响，由于两者都可以随着时间的推移而波动，所以系统的个性也可以。系统也会变得陈旧，可能会意外磨损。这些先天和后天的影响使得设计一款适合所有可能个性的电子芯片或系统变得极其困难。为了应对这一挑战，本研究将开发“芯片内统计学习”(SLIC)。SLiC是一种电子芯片/系统设计的整体方法，基于在线不断学习关键个性特征，以使系统进化到优化性能的状态。SLiC不仅将优化电子芯片的性能，还将降低成本，因为以前被认为在制造时具有弱个性的系统现在可以通过使用SLIC进行恢复。集成系统，特别是移动系统，对功率有严格的限制，这就需要从根本上重新思考如何实现学习，特别是因为系统本身既进行学习又使用由此产生的知识。因此，可以想象，一些学习任务将需要定制硬件；对于其他任务，软件甚至基于云的解决方案可能和/或必要。在这项工作中，(I)将开发用于芯片的新的学习算法，以优化每个系统级别和跨级别的操作；(Ii)将开发、设计和实施学习并适应系统个性的自我进化系统；以及(Iii)将展示使用医疗和消费电子领域的不同设计驱动器应用来提高性能和降低功耗。这项研究的优点集中在基于SLIC设计范式创建和演示一个自我进化的系统的想法。具体地说，SLIC将提供一种全面的方法，通过学习来应对系统堆栈所有级别的不确定性，从而能够在系统需求随着时间的推移而变化时在性能、功率和可靠性方面进行权衡。

Despite best efforts, electronic chips are manufactured with unique personalities that stem from the inability to precisely fabricate their underlying circuits and to create software for controlling the resulting uncertainty. It is possible to use sophisticated test methods to identify the best-performing systems, but this would result in significant cost. An electronic chip's personality is further shaped by its environment and usage, and since both can fluctuate over time, so can the system personality. Systems also grow old and can wear out unexpectedly. These nature and nurture influences make it extremely difficult to design an electronic chip or system that will operate optimally for all possible personalities. To address this challenge, this research will develop "statistical learning in-chip" (SLIC). SLIC is a holistic approach to electronic chip/system design based on continuously learning key personality traits on-line, for evolving a system to a state that optimizes performance. SLIC will not only optimize electronic chip performance but will also reduce costs since systems that were before deemed to have weak personalities at the time of fabrication can now be recovered through the use of SLIC. Integrated systems, especially mobile systems, have stringent constraints on power that necessitate a fundamental re-thinking of how to implement learning, especially since the system itself both performs the learning and uses the resulting knowledge. Therefore, it is conceivable that some learning tasks will require custom hardware; for others, software or even cloud-based solutions may be possible and/or necessary. In this work, (i) new learning algorithms for use in the chip that optimize operation at every system level and across levels will be developed; (ii) self-evolving systems that learn and adapt to the personality of a system will be developed, designed and implemented; and finally (iii) enhanced performance as well as reduced power consumption using diverse design-driver applications from the medical and consumer electronics fields will be demonstrated. The merit of this research centers on the idea of creating and demonstrating a self-evolving system based on the SLIC design paradigm. Specifically, SLIC will provide a comprehensive approach for coping with uncertainty at all levels of the system stack through learning that enables tradeoffs in performance, power, and reliability as demands on the system change over time.