Width-Aware Fine-Grained Dynamic Supply Gating: A Design Methodology for Low-Power Datapath and Memory

Abstract

With increasing contribution of leakage in total active power, run-time leakage control techniques are becoming extremely important. Supply gating provides an effective, low-overhead and technology scalable approach for active leakage reduction through the well-known "stacking effect". However, conventional supply gating approaches are typically coarse-grained in both space and time - i.e. are applied to large data path or memory blocks when an entire logic/memory block is idle for sufficiently long period. They suffer from limited applicability at run time. On the other hand, fine-grained supply gating is constrained primarily by the large wake-up delay and wake-up power overhead. In this paper, we propose a novel fine-grained width-aware dynamic supply gating (WADSG) approach to reduce both active leakage and redundant switching power in data path and embedded memory (e.g. L1/L2 cache). The approach exploits the abundance of narrow-width (NW) operands in general-purpose and embedded applications to "supply-gate" unused parts of integer execution units and memory blocks while they are in use. We introduce a novel levelized gating strategy to virtually eliminate the wake-up delay overhead. We employ the proposed WADSG approach to a super scalar processor. To reduce the wake-up power we use a width aware instruction issue policy. In case of L1 and L2 cache, we store the width information per "ways" of associative cache and supply-gate the most significant bits of the NW ways. We also propose a width-aware block allocation and replacement policy to maximize the number of NW ways. Simulation results for 45nm technology with Spec2k benchmarks show major savings (34.5%) in total processor power (considering both switching and active leakage power) with no performance impact. As a by-product, the proposed scheme also improves the thermal profile of both data path and memory.