Improving the Performance, Energy Efficiency and Security of GPUs

The work in this dissertation achieves to enhance the performance, energy-efficiency, and security of the GPUs. We noticed that, as the demand of hardware resources keeps rising in GPUs, the energy consumption becomes unaffordable and places barriers for further performance boost. To resolve this issue, we have proposed several novel GPU micro-architectures that are able to assist the GPUs to execute in an energy-efficient manner. They also provide the potential for further performance enhancement in GPUs. Firstly, we proposed a GPU register packing scheme that stores multiple narrow-width operands to a single register to save register file resources. The unoccupied registers then can be turned off to save leakage energy consumption or used to accommodate additional threads to improve the performance. Secondly, we presented a drowsy register file mechanism which can improve the energy-efficiency by putting the unused register into a low power state. Thirdly, we introduced a GPU computing units power-gating strategy by noticing the considerable idleness of the computing units. The power-gating strategy achieves a remarkable overall GPU energy reduction. Finally, in order to raise awareness about the GPU\u27s vulnerability to side-channel attacks among the researchers, we presented our research on attacking the AES algorithms that are deployed onto the GPUs and discussed appropriate countermeasures