Design Techniques for Energy-Quality Scalable Digital Systems

Energy efficiency is one of the key design goals in modern computing. Increasingly complex tasks are being executed in mobile devices and Internet of Things end-nodes, which are expected to operate for long time intervals, in the orders of months or years, with the limited energy budgets provided by small form-factor batteries. Fortunately, many of such tasks are error resilient, meaning that they can toler- ate some relaxation in the accuracy, precision or reliability of internal operations, without a significant impact on the overall output quality. The error resilience of an application may derive from a number of factors. The processing of analog sensor inputs measuring quantities from the physical world may not always require maximum precision, as the amount of information that can be extracted is limited by the presence of external noise. Outputs destined for human consumption may also contain small or occasional errors, thanks to the limited capabilities of our vision and hearing systems. Finally, some computational patterns commonly found in domains such as statistics, machine learning and operational research, naturally tend to reduce or eliminate errors. Energy-Quality (EQ) scalable digital systems systematically trade off the quality of computations with energy efficiency, by relaxing the precision, the accuracy, or the reliability of internal software and hardware components in exchange for energy reductions. This design paradigm is believed to offer one of the most promising solutions to the impelling need for low-energy computing. Despite these high expectations, the current state-of-the-art in EQ scalable design suffers from important shortcomings. First, the great majority of techniques proposed in literature focus only on processing hardware and software components. Nonetheless, for many real devices, processing contributes only to a small portion of the total energy consumption, which is dominated by other components (e.g. I/O, memory or data transfers). Second, in order to fulfill its promises and become diffused in commercial devices, EQ scalable design needs to achieve industrial level maturity. This involves moving from purely academic research based on high-level models and theoretical assumptions to engineered flows compatible with existing industry standards. Third, the time-varying nature of error tolerance, both among different applications and within a single task, should become more central in the proposed design methods. This involves designing “dynamic” systems in which the precision or reliability of operations (and consequently their energy consumption) can be dynamically tuned at runtime, rather than “static” solutions, in which the output quality is fixed at design-time. This thesis introduces several new EQ scalable design techniques for digital systems that take the previous observations into account. Besides processing, the proposed methods apply the principles of EQ scalable design also to interconnects and peripherals, which are often relevant contributors to the total energy in sensor nodes and mobile systems respectively. Regardless of the target component, the presented techniques pay special attention to the accurate evaluation of benefits and overheads deriving from EQ scalability, using industrial-level models, and on the integration with existing standard tools and protocols. Moreover, all the works presented in this thesis allow the dynamic reconfiguration of output quality and energy consumption. More specifically, the contribution of this thesis is divided in three parts. In a first body of work, the design of EQ scalable modules for processing hardware data paths is considered. Three design flows are presented, targeting different technologies and exploiting different ways to achieve EQ scalability, i.e. timing-induced errors and precision reduction. These works are inspired by previous approaches from the literature, namely Reduced-Precision Redundancy and Dynamic Accuracy Scaling, which are re-thought to make them compatible with standard Electronic Design Automation (EDA) tools and flows, providing solutions to overcome their main limitations. The second part of the thesis investigates the application of EQ scalable design to serial interconnects, which are the de facto standard for data exchanges between processing hardware and sensors. In this context, two novel bus encodings are proposed, called Approximate Differential Encoding and Serial-T0, that exploit the statistical characteristics of data produced by sensors to reduce the energy consumption on the bus at the cost of controlled data approximations. The two techniques achieve different results for data of different origins, but share the common features of allowing runtime reconfiguration of the allowed error and being compatible with standard serial bus protocols. Finally, the last part of the manuscript is devoted to the application of EQ scalable design principles to displays, which are often among the most energy- hungry components in mobile systems. The two proposals in this context leverage the emissive nature of Organic Light-Emitting Diode (OLED) displays to save energy by altering the displayed image, thus inducing an output quality reduction that depends on the amount of such alteration. The first technique implements an image-adaptive form of brightness scaling, whose outputs are optimized in terms of balance between power consumption and similarity with the input. The second approach achieves concurrent power reduction and image enhancement, by means of an adaptive polynomial transformation. Both solutions focus on minimizing the overheads associated with a real-time implementation of the transformations in software or hardware, so that these do not offset the savings in the display. For each of these three topics, results show that the aforementioned goal of building EQ scalable systems compatible with existing best practices and mature for being integrated in commercial devices can be effectively achieved. Moreover, they also show that very simple and similar principles can be applied to design EQ scalable versions of different system components (processing, peripherals and I/O), and to equip these components with knobs for the runtime reconfiguration of the energy versus quality tradeoff.