About

Abstract

The Approximate Computing for Power and Energy Optimisation ETN will train 15 ESRs to tackle the challenges of future embedded and high-performance computing energy efficiency by using disruptive methodologies.

Following the current trend, by 2040 computers will need more electricity than the world energy resources can generate. On the communications side, energy consumption in mobile broadband networks is comparable to datacenters. To make things worse, Internet-of-Things will soon connect up to 50 billion devices through wireless networks to the cloud. APROPOS aims at decreasing energy consumption in both distributed computing and communications for cloud-based cyber-physical systems. We propose adaptive Approximate Computing to optimize energy-accuracy trade-offs.

Luckily, in many parts of the global data acquisition, transfer, computation, and storage systems there exists the possibility to trade off accuracy to less power and less time consumed. As examples, numerous sensors are measuring noisy or inexact inputs; the algorithms processing the acquired signals can be stochastic; the applications using the data may be satisfied with an “acceptable” accuracy instead of exact and absolutely correct results; the system may be resilient against occasional errors; and a coarse classification may be enough for a data mining system. By introducing a new dimension, accuracy, to the design optimization, the energy efficiency can even be improved by a factor of 10x-50x. We will train the spearheads of the future generation to cope with the technologies, methodologies, and tools for successfully applying Approximate Computing to power and energy saving.

The training, in this first ever ITN addressing approximate computing, is to a large extent done by researching energy- accuracy trade-offs on circuit, architecture, software, and system-level solutions, bringing together world leading experts from European organizations to train the ESR fellows.

Why Approximate Computing?

Following the current trend, by 2040 computers will need more electricity than the world energy resources can generate [1]. Already by 2025, data centers alone will consume 20% of all available electricity [2]. A similar trend exists on the communications side where, for example, energy consumption in mobile broadband networks and mobile terminals is comparable to datacenters. In addition to the traditional personal communications, the Internet-of-Things (IoT) will soon connect up to 50 billion devices [3] through wireless networks to the cloud, which will accelerate these trends. In order to alleviate the energy issues, APROPOS will contribute to decreasing energy consumption in both distributed computing and communications for cloud-based cyber-physical systems by introducing an adaptive energy-aware approximate computing overlay. Approximate and transprecision (i.e., adaptive precision) computing paradigms combined with application-specific processing structures are the key elements in achieving the required energy efficiency improvements. Since energy consumption is the product of (computing or communication) time and average power consumption of the device while carrying out an operation, these two factors, time and power, must be addressed for achieving energy savings. The precious and novel third dimension, accuracy adjustment for decreasing time and power, is the key contribution of the APROPOS network.

Why now?

Despite the recent advances in semiconductor technology and energy-aware system design, the overall energy consumption of computing and communication systems is rapidly growing. On one hand, the pervasiveness of these technologies everywhere in the form of mobile devices, cyber-physical embedded systems, sensor networks, wearables, social media and context-awareness, intelligent machines, broadband cellular networks, cloud computing, and IoT has drastically increased the demand for computing and communications. On the other hand, the user expectations on features and battery life of on-line devices are increasing all the time, creating another incentive for finding good trade-offs between performance and energy consumption. For instance, IoT is rapidly extending to various application areas, including health technology, smart homes, smart cities, intelligent transport systems, connected cars, and machine-to-machine communication. IoT as a trend is tightly related to Big Data analytics, cloud computing, and to the emerging 5G communications infrastructure. IoT sensor nodes are responsible for collecting the data for analysis, machine learning (ML) and long-term storage and deployment in the cloud.

However, cloud computing has introduced some drawbacks compared to local processing of data, in particular high latencies and energy consumption, due to the communication overheads and inefficient general-purpose processing engines. A recent trend is to bring the computations closer to the endpoint devices by applying Edge or Fog Computing at the edge of the network [5], with the advantages of lower total energy, lower latency, reduced communication needs, and location-awareness. Edge devices usually have a limited amount of resources: computational power, memory, and energy storage. These limitations have an impact on the capability of those devices to process data locally. In addition to cloud computing, also the mobile broadband network is a major energy consumer. 5G networks will support energy saving when data transfers can be minimized, in a manner resembling the stop-and-start functionality of modern car engines. The main challenge there is to allow for such energy-saving periods in data transfers, and to render the transfer bursts as low-energy as possible. Also, the (IoT and mobile) devices and communications in the periphery of 5G networks need to be optimized for lower energy consumption. The battery-operated devices are the most dependable on low energy dissipation, although not that big contributor to energy consumption in the global scale.

To improve the energy efficiency of computing and communications, radical efforts are needed. As pointed out, the energy consumption is the product of time and power. We postulate that, by allowing a decrease in the accuracy of processing at different parts of the computing chain, we can achieve less power consumption and shorter computing time. Trading off accuracy for improved power consumption and/or performance is generally known as Approximate Computing (AC). Transprecision Computing (TC) is a subset and extension of AC, where adaptive precision is applied to different parts of the computation, possibly without affecting the end-to-end accuracy of the system. There have been a number of baseline solutions in international research for AC.

Why us?

The beneficiaries are European top players in AC and TC, most of them with experience in European, transnational and national projects on different aspects of these technologies. The supervisors have also excellent track records in training PhDs, carrying out and leading research, and disseminating the results in books, journals and conferences (see section 1.3. Quality of Supervision). The partners are innovative research-oriented companies, eager to contribute to training young talents on the subject. The coordinator has prior experience on running a successful ITN project [7] as well as on organizing seasonal schools, workshops, and international events. By now, there has been no MSCA ITN addressing Approximate and Transprecision Computing, and this consortium is definitely the one to take on such a job, to train skilled and competent ESRs for the needs of Europe to develop greener computing and communication solutions. This is especially needed in order to bridge the gap to the US, where AC has recently gained a lot of attention, and place Europe in the forefront of this emerging area.

As concluded in a recent survey by Xu et al. [8]: “existing work in approximate computing has shown great promises for certain domains, but significant innovations and research efforts are needed to enable approximate computing as a practical mainstream computing paradigm.” This is exactly what we are aiming for in this project, basing our work on the beneficiaries’ own research and the other state-of-the-art, bringing the Early Stage Researchers (ESRs) to the world’s top-level knowledge and expertise on approximate computing, and introducing AC to various applications by means of the research projects of the ESRs recruited. The application areas are ranging from multimedia, virtual reality and communications to cryptography/security and scientific computations. In the APROPOS project, approximate and transprecision computing are applied to computing and digital communications widely in a cloud-based distributed system framework. This is done by researching trade-offs on circuit, architecture, software and system-level solutions, for saving resources (circuit area, memory), time, power and energy while reaching acceptable levels of accuracy or precision

[1] “By 2040, computers will need more electricity than the world can generate,” The Register, July 25, 2016. www.theregister.co.uk, accessed on September 27, 2019.

[2] João Marques Lima, “Data centres of the world will consume 1/5 of Earth’s power by 2025,” Data Economy, December 12, 2017, data-https://economy.com/data-centres-world-will-consume-1-5-earths-power-2025/, accessed on December 5, 2019.

[3] “IoT Connections to Grow 140% to Hit 50 Billion By 2022, As Edge Computing Accelerates ROI,” Juniper Research, June 12, 2018, https://www.juniperresearch.com/press/press-releases/iot-connections-to-grow-140-to-hit-50-billion

[4] “Rebooting the IT Revolution,” Semiconductor Research Corporation, Sept. 2015, https://www.src.org/newsroom/rebooting-the-it-revolution.pdf

[5] Flavio Bonomi, Rodolfo Milito, Jiang Zhu, and Sateesh Addepalli, ”Fog Computing and Its Role in the Internet of Things,” in Proc. ACM SIGCOMM, Aug. 17, 2012, Helsinki, Finland, pp. 13-16.

[6] Amir Vahid Dastjerdi and Rajkumar Buyya, “Fog Computing: Helping the Internet of Things Realize Its Potential,” IEEE Computer, 49(8), 2016, pp. 112-116.

[7] Jari Nurmi and Elena-Simona Lohan, “MULTI-POS: Marie Curie Network in Multi-technology Positioning,” in Proc. Design, Automation & Test in Europe (DATE 2016), Dresden, Germany, March 14-18, 2016.

[8] Qiang Xu, Todd Mytkowicz, and Nam Sung Kim, “Approximate computing: A survey,” IEEE Design & Test, vol. 33, pp. 8–22, Feb 2016.