Wake-up logic is responsible for informing instructions in the Window that are waiting to execute, about the availability of their input operands. The conventional method of wake-up consumes a significant percentage of the Instruction Window energy. Reducing the wake-up energy also addresses the Instruction Window hot spot problem caused due to the high power density of the Instruction Window. In this work, we investigate the energy and power savings…
Contents
Introduction
1.1 Motivation
1.2 Contributions of this thesis
1.3 Organization of the Thesis
2 Background
2.1 Out-of-Order (OOO) Execution .
2.2 The Instruction/Execution pipeline .
2.3 The Instruction Window
2.4 Related Work
3 DL-based Wake-up
3.1 Motivation and Outline
3.2 Theory and Implementation
3.3 A Comparison with the N-use scheme
3.4 Impact on Performance (IPC)
3.5 Impact of Clock Cycle
3.6 Impact on Area/Space
3.7 Branch Misprediction/Instruction Squashes
4 Analysis of Energy Consumption
4.1 Energy Consumed in DL-based Wake-up
4.2 Energy Consumed in Conventional Wake-up
4.2.1 Tag Drive Energy
4.2.2 Tag Match Energy
4.2.3 Match-OR Energy
4.3 A Comparison of the Energy Consumed in the DL-based Scheme and the Conventional Scheme
4.4 Effect of Dlist Length on Wake-up Energy
4.5 Scalability of DL-based Scheme with Instruction Window Size and Issue Width
5 Need-Based DL (NBDL) Wake-up Scheme
5.1 Motivation
5.2 Parent Instructions
5.3 Implementation
5.4 Energy Analysis
5.5 Effect of NBDL scheme on delay
5.6 Impact of NBDL scheme on Processor Area
6 Experimental Analysis
6.1 Experimental Setup
6.1.1 Simulator
6.1.2 Benchmark
6.1.3 Baseline
6.1.4 Microarchitectural parameters
6.1.5 Simulation Parameters
6.2 Results
6.2.1 Performance Results
6.2.2 Energy Savings in Wake-up
6.2.3 Instruction Window Power Savings
6.2.4 Energy Delay Product
6.2.5 Scalability with Increasing Window Size and Issue Width
6.2.6 Performance of DL-based and NBDL Schemes for a Split Window Organization
6.2.7 Overall Processor Power Savings
Bibliography
Author: Kakaraparthi, Himabindu
Source: University of Maryland
Download URL 2: Visit Now