Reducing aging on scratchpad memory using temporal- and FSM-based power management

Abstract

Scratchpad memory (SPM) which is software-controlled SRAM has lower power consumption compared to cache. It is becoming widely used in CPU, GPU and embedded devices. Meanwhile, as CMOS technology continues shrinking, negative bias temperature instability (NBTI) has become a major reliability concern. Recent studies show that SRAM suffers from NBTI effect, so it is a significant issue for SPM. While most research has focused on mitigating aging on cache, little attention are focusing on SPM aging mitigation. In this paper, we propose a temporal-based power management technique to turn SPM pages into sleep modes according to its recently used history. To account for different access behaviors of different applications, we propose a dynamic threshold adjustment algorithm to adjust a predefined threshold (P th ) to turn off the idle pages dynamically. We also propose a FSM-based power management technique that combines both voltage scaling and power gating techniques to change the states of SPM pages into drowsy mode or sleep mode at runtime. The experimental results show that degradation is improved more than 7% when both techniques are used simultaneously.