Exploration of Energy Efficient Memory Organizations Exploiting Data Variable Based System Scenarios

Modern embedded systems are capable of performing a wide range of tasks and their popularity is increasing in many different application domains. The recent progress in semiconductor processing technology greatly improves the performance of the embedded systems, due to the increased number of transistors on a single chip. The continuous performance improvement, for example higher clock frequencies and lower supply voltage, increases the possibilities for new embedded system designs. However, many embedded systems rely on a battery source, which puts a significant limitation on their lifetime and usage. The management of the energy consumption is a key factor towards increasing the lifetime of an embedded system relying on a battery source. A typical embedded system includes one or more processing elements (CPUs, GPUs, embedded processors etc.), a memory subsystem (cache, scratchpad, flash, etc.) and application specific hardware (antennas, sensors, etc.). The memory subsystem contributes significantly to the overall energy consumption as shown in many studies of embedded system applications and platforms. Especially for applications that are data intensive, the memory architecture may have the highest energy footprint of the whole embedded system. A hardware/software co-design methodology is proposed for the reduction of the energy consumption in the memory subsystem based on the system scenario methodology. In general, system scenario methodologies propose the use of different platform configurations in order to exploit runtime variations in application needs. The current methodology exploits variations in memory needs during the lifetime of an application in order to optimize energy usage. The different resource requirements, that change dynamically at run-time, are grouped into scenarios to efficiently handle a very large exploration space. Apart from the development of the methodology, an extended memory model is included in this work. The memory models is based on existing state-of-the-art memories, available from industry and academia. In addition, the impact of the technology scaling is studied and the effectiveness of the proposed methodology is analyzed for the future memory architectures. We also investigate the combination of the developed methodology with known code transformation techniques, specifically data interleaving. The proposed design methodology aims at being compatible with the already available code optimization techniques. Actual decrease in memory energy consumption for a selection of studied applications ranges between 30% and 60%