Energy-Efficient Programmable Accelerators

50年来，计算系统一直受益于按美元计算的成倍增长的性能。1971年，第一个微处理器是用10微米晶体管制造的，而最近的设计是用14纳米晶体管制造的。由此产生的晶体管密度提高了5个数量级，时钟频率提高了3个数量级，使计算从只处理简单的商业计算转变为支持在精度上与人类匹敌的机器学习算法，并导致了有望广泛提高便利性和生产率的新兴“物联网”。*然而，在5 nm工艺节点威胁到这种未来愿景后，晶体管缩放的速度可能会显著放缓。事实上，晶体管阈值电压的调整在十年前就基本停止了，导致时钟频率停滞不前，转向多核，以及最近对“暗硅”的担忧。虽然替代技术(例如量子计算)可能提供优秀的长期解决方案，但这些技术将需要数十年才能完全开发出来，从而导致“休闲期”，在此期间，除非找到替代技术，否则计算系统的能力将停滞不前。原则上，通过开发专门的硬件，计算系统的能力可以提高几个数量级。这一承诺是以降低灵活性和/或困难的编程模型为代价的。*这项研究计划的长期目标是使不同软件应用程序的每美元计算能力与当今的计算硬件相比有数量级的改进。这样的改善几乎肯定会使企业和整个社会受益。例如，这样的增长可能会造福社会，使更复杂的机器学习能够应用于嵌入式物联网设备。计算能力的数量级增长可能会转化为两种冰箱之间的区别：一种是支持物联网的冰箱，它会在牛奶即将在约会前到达时向你发送电子邮件；另一种是识别你对食物的选择感到厌倦，并根据过去的偏好和你家人的健康目标建议购物清单的冰箱。*这项研究计划将通过探索软件和硬件方法来解决这些目标，以提高软件开发的简便性和通过利用专门的硬件提高计算能力之间的权衡。预期的结果是洞察如何最好地构建未来的计算系统，使目前强大的软件行业在摩尔定律结束后仍能生存。加拿大将从中受益，因为有越来越多的计算机硬件(英特尔、AMD、高通)和软件(艺电、IBM、微软)公司在加拿大设有研发中心，可以随时利用这些洞察力。**

For 50 years computing systems have benefited from exponentially increasing performance per dollar. The first microprocessor was fabricated using 10 micrometer transistors in 1971 while recent designs have been manufactured using 14 nanometer transistors. The resulting five-orders of magnitude increase in transistor density combined with three orders of magnitude increase in clock frequency has enabled computing to change from handling only simple business calculations to supporting machine learning algorithms rivaling humans in accuracy and led to the emerging "internet of things" that promises to broadly increase convenience and productivity. ***However, the rate of transistor scaling will likely slow significantly after the 5nm process node threatening such visions for the future. Indeed, transistor threshold voltage scaling essentially stopped a decade ago resulting in stagnant clock frequencies, a shift to multicore and more recent concerns about "dark silicon". While alternative technologies (e.g., quantum computing) may provide excellent long term solutions these will take decades to fully develop resulting in a "fallow period" during which computing system capability will stagnate unless alternatives are found. In principle, computing system capability can be improved by orders of magnitude by exploiting specialized hardware. This promise comes at the expense of reduced flexibility and/or difficult programming models. ***The long-term goal of this research program is to enable order-of-magnitude improvements in computing capability per dollar versus today's computing hardware for diverse software applications. Such an improvement will almost certainly benefit business and society in general. For example, such increases could benefit society enabling more sophisticated machine learning to be applied embedded internet-of-things devices. An order of magnitude gain in computing capability could translate into the difference between a IoT enabled fridge that emails you if the milk is about to reach its use before date, and one that recognizes you are getting bored with your food choices and suggests a shopping list based upon past preferences and your family's health goals.***This research program will tackle these goals by exploring both software and hardware approaches to improve the tradeoff between ease of software development and increasing computing capability by exploiting specialized hardware. The expected outcome is insights into how best to structure future computing systems so that the currently robust software industry remains viable past the end of Moore's Law. Canada will benefit because there is a growing number of computer hardware (Intel, AMD, Qualcomm) and software (Electronic Arts, IBM, Microsoft) companies with an R&D presence here that can readily leverage these insights.**