Exploiting Dark Silicon for Energy Efficiency EXECUTIVE SUMMARY

Application datasets grow faster than Moore’s Law [7,8], both in personal and desktop computing, as well as in the scientific domain. For example, during March 2011, over 1.6PB of data were gen-erated and transferred for processing among the Tier-1 sites of CERN’s Large Hadron Collider computing grid [3]. Processing all these data requires unprecedented amounts of computation, with commensurate energy demands. Computing’s insatiable appetite for energy has created an IT industry with approximately the same carbon footprint as the airline industry, accounting for 2% of the global greenhouse gas emissions. Computers consumed 408 TWh globally in 2010, and with a forecasted 10 % annual growth on installed computing systems, this figure will rise, nega-tively impacting the environment and the economics of computing [1]. Unfortunately, a large frac-tion of this energy is wasted. Modern processors account for 38 % of the power consumption of a computing system [5], but the general-purpose computing substrate they provide is hugely power inefficient. For example, compared to an optimized ASIC design, conventional multicores con-sume 157–707 times more energy when executing various phases of an H.264 encoder [6]. We postulate that a promising path towards energy-efficient data processing is the development of multicore processors with specialized cores. To implement these cores, we propose to utilize sili-con that would otherwise remain powered-off, or “dark”, due to the physical constraints of proces-sor chips. Specialized multicores can leverage the underutilized die area to overcome the power barrier, delivering significantly higher performance for the same power envelope, thereby signifi-cantly reducing energy consumption. In the remaining of the paper we describe these constraints and how they give rise to dark silicon, explain why utilizing dark silicon is a viable solution, esti-mate the possible energy benefits that could be achieved by specialized computing on commercial workloads (e.g., online transactional processing, decision support, and web server workloads), and conclude with a set of research directions for the software community