Topology specialization for networks-on-chip in the dark silicon era

Modarressi, M ; Sharif University of Technology | 2018

262 Viewed
  1. Type of Document: Article
  2. DOI: 10.1016/bs.adcom.2018.03.009
  3. Publisher: Academic Press Inc , 2018
  4. Abstract:
  5. Following Moore's law, the number of transistors on chip has grown exponentially for decades. This growing transistor count, coupled with recent architecture and compiler advances, has resulted in an unprecedented exponential performance increase of computers. With the end of Dennard scaling, however, the power required to operate all transistors at the full performance level simultaneously grows across the technology generations. Consequently, chips will keep an increasing fraction of transistors power gated or dark to remain within the power envelope. The power-gated part of the chip, known as dark silicon, is expected to comprise a significant portion of the die real estate in new technology generations. In addition to power limitations, the limited and nonscalable off-chip bandwidth is the second source of dark silicon. The key challenge to improving performance in the dark silicon era, consequently, is how to efficiently leverage transistors when they cannot all be powered at the same time. Core specialization shows promise in addressing this challenge by improving the performance and energy efficiency of applications. This method leverages the dark silicon to build many diverse power-efficient accelerators and each application only activates a subset of cores that best match its processing requirements. In this chapter, we propose a network-on-chip specialization method that leverages dark routers of a partially active many-core chip to customize the topology for active cores, effectively reducing the power consumption and latency of communication. © 2018 Elsevier Inc
  6. Keywords:
  7. Dark silicon ; Network-on-chip ; Reconfiguration ; Topology
  8. Source: Advances in Computers ; Volume 110 , 2018 , Pages 217-258 ; 00652458 (ISSN); 9780128153581 (ISBN)
  9. URL: https://www.sciencedirect.com/science/article/pii/S0065245818300226