Computer architecture design is in a new era where performance is increased by replicating processing cores on a chip rather than making CPUs larger and faster. This design strategy is motivated by the superior energy efficiency of the multi-core architecture compared to the traditional monolithic CPU. If the trend continues as expected, the number of cores on a chip is predicted to grow exponentially over time as the density of transistors on a die increases. A major challenge to the efficiency of multi-core chips is the energy used for communication among cores over a Network on Chip (NoC). As the number of cores increases, this energy also increases, imposing serious constraints on design and performance of both applications and architectures. Therefore, understanding the impact of different design choices on NoC power and energy consumption is crucial to the success of the multi- and many-core designs. This dissertation proposes methods for modeling and optimizing energy consumption in multi- and many-core chips, with special focus on the energy used for communication on the NoC. We present a number of tools and models to optimize energy consumption and model its scaling behavior as the number of cores increases. We use synthetic traffic patterns and full system simulations to test and validate our methods. Finally, we take a step back and look at the evolution of computer hardware in the last 40 years and, using a scaling theory from biology, present a predictive theory for power-performance scaling in microprocessor systems.
multi-core, many-core, energy consumption, communicaiton locality, scaling
Level of Degree
Department of Computer Science
First Committee Member (Chair)
Second Committee Member
Third Committee Member
Bezerra, George. "Energy consumption in networks on chip : efficiency and scaling." (2012). http://digitalrepository.unm.edu/cs_etds/21