Energy Efficiency of Micropipelines Under Wide Dynamic Supply Voltages Abstract— The contradictions among high performance, low power and unpredictable energy supply motivate a large part of current mobile and embedded system design research especially on systems that can work under a wide range of voltages. Concurrency has been used to improve performance and/or efficiency, but has not been properly studied with widely variable supply voltages. In this paper, a self-timed micropipeline is designed to investigate system behaviour under a wide range of voltages, focusing on asynchrony and the relationship between the degree of concurrency and all important performance metrics including the amount of computation. Results suggest that above threshold, the amount of computation per given amount of energy is practically insensitive to the degree of concurrency, but below threshold the dependency on the degree of concurrency goes up significantly. This indicates that probably the best architectures for energy efficiency are asynchronous data-flow ones and those operating in sub-threshold. Keywords—High performance; Efficient system; Ultra low power; Micropipeline; Parallism; Asynchrony; Adaptive; Self-timed circuit; I. INTRODUCTION CMOS electronic circuit designs have been driven by industrial needs towards several extremes, such as high performance and low power. On entering the deep sub-micro (nanometer) technology era, the large numbers of functions that can be integrated provide possibilities for extensive concurrency. However, [1] pointed out that the conventional method of packing as many function blocks as possible with power management could lead to high power consumption and heat, which will slow down the circuitry [2], causing a vicious circle to develop. Therefore low power became the most important concern [3]. On the other hand, in portable devices powered by batteries or other alternatives such as harvesting power sources [4], power efficiency is the most important aspect in order to extend the life time of devices and to execute more tasks. Voltage is a key parameter related to power because of their quadratic relationship, leading to a number of power optimization techniques, such as DVS, DVFS, etc. Dropping the VDD causes a performance reduction with current technologies. As a result, performance and power consumption should be considered together in a variable voltage environment caused by both active voltage scaling and potential supply uncertainties. One of the ‘knobs’ that a power and performance optimization scheme can tune is the degree of concurrency, given the prominence of highly parallel systems. Recently, a novel power supply which is able to provide a wide band of voltage levels has been proposed [5]. And some preliminary work has already shown benefits from the wide range working voltages [6]. Tuning the power-performance balance using the degree of concurrency is still in its infancy [7] and a wide band of working voltages has not been comprehensively studied, especially on highly concurrent systems, such as multi-core systems and pipeline structures. In this paper, a study of data computation efficiency under a wide range of dynamic voltages is presented. The authors of this paper are trying to shed more light on the relations between the various dynamic parameters of circuits and systems. In the other words, the authors would like to investigate power efficiency, and circuit/system reliabilities on this very popular highly parallel mechanism under a wide band of voltage level. The results are expected to benefit to designing highly parallel multi-core systems. The main contributions of the paper include designing experiments for the exploration of the performance and power metrics of concurrent systems using a highly parallel micropipeline, and experiment result analysis leading to an interesting conclusion. In addition, circuits such as a C element with set/reset functions are designed to support the experiment. The remainder of this paper is organized as follows: Section 2 briefly introduces micropipelines and their properties; Section 3 describes the actual design of the experimental system including the data injection circuits; In section 4 the experimental results are analyzed; and Section 5 concludes the study and discusses future work. II. MICROPIPELINE Asynchronous systems will play an important role in the future [8], because of its potential advantages low power, average performance, process variation tolerance, etc. The micropipeline [9] is a well-known asynchronous mechanism. It can be used in both data path and control, unlike conventional pipelines [10] which are used in data path. However, one of its potential uses, to study how the degree of concurrency affects system behavior, has not been explored. Compared with the synchronous pipeline [10] which contains a fixed number of data items, data items can be injected into a Abdullah Baz, Delong Shang, Fei Xia, Xuan Gu, and Alex Yakovlev uSystem Group, School of Electrical & Electronic Engineering, Newcastle University, U.K. {delong.shang,Abdullah.baz,fei.xia,x.gu3,alex.yakovlev}@ncl.ac.uk