Energy-Efficient Instruction Streams for Embedded Processors

The information and communication technology (ICT) sector is consuming an increasing proportion of global electricity production. At the heart of ICT are computer systems that perform computations on data. During the last two decades the rate of improvement in their energy-efficiency has declined.

Although the best energy-efficiency is obtained with systems that execute a fixed set of tasks, they are inflexible. The flexibility can be improved with programmability in the form of instruction set support. However, the resulting instruction stream incurs area and energy overheads. To reduce them, a variety of instruction stream components have been proposed. In particular, first-level components closest to the compute elements in memory hierarchy have been studied carefully. Although dynamically controlled components are easy to integrate into systems and allow a simplified view for programmers, software-controlled components have been shown to improve energy-efficiency. While they have been researched previously, the energy-efficiency of programmable processors is not yet at the level of fixed function accelerators, and there is room for novel methods utilizing fine-grained software control.

In addition to the novel first-level components, memory technologies used in the instruction streams have received wide interest in recent years. While emerging memory technologies may result in extreme efficiency in future devices, customizing the currently used technologies may offer benefits with less effort for the near future.

This thesis proposes methods to improve the instruction stream energy-efficiency with the goal of reducing the gap between instruction set programmable and fixed-function computer systems. By evaluating design choices in first-level instruction stream components and allowing fine-grained software control, energy consumption is reduced. Similarly, efficiency of unconventional instruction memories is improved by exposing their contents to the program compiler. These methods help future computer systems to keep increasing their energy-efficiency and reduce the proportion of ICT from global energy consumption.