Design of an Online Data Clustering Method for Multi-bank Scratchpad Memory

  • 林 弋喬

Student thesis: Master's Thesis


Power consumption of on-chip memory is divided two parts dynamic power and leakage power With the scaling of CMOS devices the portion of leakage power consumption increases and it can go over 50% of total power consumption in a CMOS device Scratchpad memory (SPM) a software-controlled on-chip memory has less access energy and higher area density when compared to an ordinary cache Multi-bank SPM was proposed to further reduce the energy consumption This paper proposed OCBAS a novel Online Data Clustering for Multi-bank Scratchpad memory OCBAS identifies the hotness of each SPM page and clusters pages with different degrees of hotness into different SPM banks increasing idleness for cold SPM banks and hence reducing the leakage energy of the SPM Moreover OCBAS allows cold SPM banks to enter the low power mode more aggressively further reducing the leakage energy Offline profiling is not required in OCBAS The evaluation results show that OCBAS can reduce the energy delay product (EDP) by up to 37 71% (18 09% on average) in 32KB 4-bank 25 33% (15 71% on average) in 32KB 8-bank 28 99% (13 45% on average) in 16KB 4-bank 30 18% (10 05% on average) in 16KB 8-bank SPM compared to conventional time-out-based policy Moreover the area overhead of the hardware support is insignificant (about 0 72%)
Date of Award2016 Feb 2
Original languageEnglish
SupervisorDa-Wei Chang (Supervisor)

Cite this