Energy Reduction Techniques for Caches and Multiprocessors

E nergy consumption is a growing concern in many areas of computer architecture. Not only for the handheld embedded market, but also for desktop machines and high-end server facilities, there is a demand for ever increasing processing power while maintaining or even decreasing energy consumption. For processors embedded in battery-powered devices, consumers both demand an increasing number of features and an increase of battery lifetime. For commodity desktop and high-end server systems, the demand to reduce energy consumption is mostly fueled by cost, environmental issues, and the wish to have systems without noisy cooling systems. This dissertation studies several techniques that aim at reducing energy consumption in processors. Part of the techniques presented in this dissertation focusses at reducing energy consumption by decreasing the amount of data transferred between a processor and external memory. Since memory is one of the known bottlenecks in computer systems, manufacturers had to employ increasingly aggressive techniques in the past decades to increase performance. The techniques proposed in this dissertation target at improving or at least maintaining performance, while reducing the amount of energy dissipated in the memory subsystem. Another part of this dissertation focusses on reducing energy by lowering the speed of nodes in multiprocessor systems in combination with turning off some of these nodes. Multiprocessor systems have gained significant interest in the past years, mostly because power constraints have prevented further increasing clock frequencies and because instruction level parallelism has suffered from diminishing returns. Due to the way how energy is dissipated in semiconductor fabric, using multiple cores on a reduced frequency is an effective way to reduce energy consumption. Due to decreasing sizes of the components from which processors are built, it is expected that this energy model will change significantly in future years. Some of the techniques presented in this dissertation aim at reducing energy consumption in such contemporary and near-future multiprocessor systems.