Date of Award
Master of Science
Electrical and Computer Engineering
Arun K. Somani
With the ever-decreasing feature sizes, static power dissipation has become a concern in computing devices. On-chip memories are a major contributor towards the processor’s leakage power dissipation due to their large transistor count. We propose a Low Leakage Energy Architecture for Caches, called LowLEAC to minimize the static power dissipation in caches made of CMOS SRAM cells. This technique is based on keeping only k most recently used cache lines powered on other lines powered off to reduce the leakage power dissipation. The control however increases the dynamic power due to re-fetching of data. To overcome that, we deploy CMOS compatible nonvolatile SRAM cell, called cNVSRAM, to implement caches. The cNVSRAM cell works as a conventional SRAM in the regular mode and saves the data in a non-volatile back up when a cache line is turned off or put in the sleep mode. The non-volatile back up mode helps improve the dependability of the cache and avoids the penalty occurred due to loss of data from the inactive cache lines. With a small area penalty, LowLEAC achieves 18% energy savings with insignificant impact on the performance. LowLEAC is a suitable architecture for cache memory in mobile computing devices to minimize battery power consumption and reduce heat.
Rashmi Parisa Girmal
Girmal, Rashmi Parisa, "LowLEAC: Low leakage energy architecture for caches" (2017). Graduate Theses and Dissertations. 15308.