Multiple sleep modes leakage control in peripheral circuits of a all major SRAM-based processor units

  • Authors:
  • Houman Homayoun;Avesta Sasan;Aseem Gupta;Alex Veidenbaum;Fadi Kurdahi;Nikil Dutt

  • Affiliations:
  • University of California Irvine, Irvine, CA, USA;UC Irvine, Irvine, CA, USA;UC Irvine, Irvine, CA, USA;UC Irvine, Irvine, CA, USA;UC Ircine, Irvine, CA, USA;UC Irvine, Irvine, CA, USA

  • Venue:
  • Proceedings of the 7th ACM international conference on Computing frontiers
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Leakage currents in on-chip SRAMs: caches, branch predictor, register files and TLBs, are major contributors to the energy dissipated by processors in deep sub-micron technologies. High leakage also increases chip temperature and some SRAM-based structures become thermal hotspots. Previous work has addressed major sources of SRAM leakage in memory cells and bit-lines, making remaining SRAM components, in particular large drivers, the primary source of leakage. This paper proposes an approach to reduce this source of leakage in all major SRAM-based units of the processor, controlling them in a uniform way, yet treating each unit individually based on its behavior and memory organization. The new approach uses multiple bias voltages in sleep transistors allowing a trade-off between leakage reduction and wakeup delay in multi-stage peripheral drivers. Four low-power modes are defined, from basic to ultra low power, and SRAMs dynamically transition between these modes to minimize leakage without sacrificing performance. A novel control mechanism monitors and predicts future processor behavior for mode control. The leakage reduction in individual units is evaluated and shown to vary from 25% for IL1 to 75% for L2 caches. Resulting temperature reduction, including the effect of positive feedback between temperature and leakage power, is evaluated. A significant temperature reduction is achieved in each unit. It is also shown to reduce hot spots in the instruction TLB and branch predictor.