Multi-criteria optimization for energy-efficient multi-core systems-on-chip

The steady down-scaling of transistor dimensions has made possible the evolutionary progress leading to today’s high-performance multi-GHz microprocessors and core based System-on-Chip (SoC) that offer superior performance, dramatically reduced cost per function, and much-reduced physical size compared to their predecessors. On the negative side, this rapid scaling however also translates to high power densities, higher operating temperatures and reduced reliability making it imperative to address design issues that have cropped up in its wake. In particular, the aggressive physical miniaturization have increased CMOS fault sensitivity to the extent that many reliability constraints pose threat to the device normal operation and accelerate the onset of wearout-based failures. Among various wearout-based failure mechanisms, Negative biased temperature instability (NBTI) has been recognized as the most critical source of device aging. The urge of reliable, low-power circuits is driving the EDA community to develop new design techniques, circuit solutions, algorithms, and software, that can address these critical issues. Unfortunately, this challenge is complicated by the fact that power and reliability are known to be intrinsically conflicting metrics: traditional solutions to improve reliability such as redundancy, increase of voltage levels, and up-sizing of critical devices do contrast with traditional low-power solutions, which rely on compact architectures, scaled supply voltages, and small devices. This dissertation focuses on methodologies to bridge this gap and establishes an important link between low-power solutions and aging effects. More specifically, we proposed new architectural solutions based on power management strategies to enable the design of low-power, aging aware cache memories. Cache memories are one of the most critical components for warranting reliable and timely operation. However, they are also more susceptible to aging effects. Due to symmetric structure of a memory cell, aging occurs regardless of the fact that a cell (or word) is accessed or not. Moreover, aging is a worst-case matric and line with worst-case access pattern determines the aging of the entire cache. In order to stop the aging of a memory cell, it must be put into a proper idle state when a cell (or word) is not accessed which require proper management of the idleness of each atomic unit of power management. We have proposed several reliability management techniques based on the idea of cache partitioning to alleviate NBTI-induced aging and obtain joint energy and lifetime benefits. We introduce graceful degradation mechanism which allows different cache blocks into which a cache is partitioned to age at different rates. This implies that various sub-blocks become unreliable at different times, and the cache keeps functioning with reduced efficiency. We extended the capabilities of this architecture by integrating the concept of reconfigurable caches to maintain the performance of the cache throughout its lifetime. By this strategy, whenever a block becomes unreliable, the remaining cache is reconfigured to work as a smaller size cache with only a marginal degradation of performance. Many mission-critical applications require guaranteed lifetime of their operations and therefore the hardware implementing their functionality. Such constraints are usually enforced by means of various reliability enhancing solutions mostly based on redundancy which are not energy-friendly. In our work, we have proposed a novel cache architecture in which a smart use of cache partitions for redundancy allows us to obtain cache that meet a desired lifetime target with minimal energy consumption.