SHF:Small:Collaborative Research: Elastic Fidelity: Trading-off Computational Accuracy for Energy Efficiency

点击翻译按钮获取中文摘要

Energy and power consumption have become a critical issue ranging from microarchitectures to large-scale data centers and supercomputers. Conservative estimates suggest that the information technology industry world-wide energy consumption is in excess of 400 TWh and growing, generating roughly the same carbon footprint as the airline industry, accounting for 2% of global emissions. At the same time, the power constraints of chips hamper their performance, and the shrinking transistor geometries and low supply voltages increase the severity of processor variations resulting in higher timing error rates. High error rates lead to a significant drop in yield and increased manufacturing costs, calling for designs that are able to withstand them. This project seeks to understand and explore the novel paradigm of elastic fidelity computing. Elastic fidelity computing capitalizes on the observation that many applications can naturally tolerate errors, and that not all of them need to run at 100% fidelity all the time. Specifically, the goal of this work is to understand the error models of various hardware components as they relate to data movement, storage, and computation, and simultaneously to understand the error resiliency of applications and re-architect them to leverage elastic fidelity. Elastic fidelity offers potentially transformative effects for science and society, by challenging conventional wisdom and taking a fresh look at the interplay of errors, output quality and energy efficiency for an important class of pervasive streaming and data-intensive applications. More specifically, elastic fidelity promises significant energy savings that can put computing on an environmentally sustainable path, by lowering the operational costs in major economic sectors, and making the manufacturing of future chips cheaper by relaxing the accuracy requirements of hardware components. The results of this research will be disseminated through publications, workshops, advanced curriculum, and releases of the developed infrastructure in the public domain. To accelerate broad societal effects, the project participants will seek to foster technology transfer by promoting collaboration and industry involvement through presentations and site visits.