Abstract
High-speed serial network interfaces are gaining wide use in connecting multiple processors and peripherals in modern embedded systems, thanks to their size advantage and power efficiency. Many such interfaces also support multiple data rates, and this ability is opening a new dimension in the power/performance trade-offs between communication and computation on voltage scalable embedded processors. To minimize energy consumption in these networked architectures, designers must not only perform functional partitioning but also carefully balance the speeds between communication and computation, which compete for time and energy. Minimizing communication power without considering computation may actually lead to higher energy consumption at the system level due to elongated on-time as well as lost opportunities for dynamic voltage scaling on the processors. We propose a speed selection methodology for globally optimizing the energy consumption in embedded networked architectures. We formulate a multidimensional optimization problem by modeling communication dependencies between processors and their timing budgets. This enables engineers to systematically solve the problem of optimal speed selection for global energy reduction. We demonstrate the effectiveness of our speed selection approach with an image processing application mapped onto a multi-processor architecture with a multi-speed Ethernet.