Global Communication and Memory Optimizing Transformations for Low Power Systems

In this paper we illustrate the crucial impact of memory related power consumption on the global system power budget, in particular for multi-dimensional signal processing subsystems and for table-based communication network components. In addition, several design methods are described to decrease this impact by transforming the initial signal or data processing speciication, especially in terms of control ow transformations. Several of these methods are becoming supported by means of computer-aided design techniques. 1 Low power system design In practice, there are many ways to realize a given application. The system designer has for instance the choice between many algorithmic procedures for a desired behaviour (e.g. sorting). In order to make a motivated choice, it should be based on "quan-tiiable cost" measures (number of components, pin count, power consumption, area of custom components), augmented with other issues like market demand , time-to-money and competition. In general, a good trade-oo between these characteristics is crucial so there is a need for fast and early feedback already at the algorithm level without always descending to detailed RT/logic realization or layout level. It is also expected that only a relative comparison is required so that an accuracy of the measures within a few dozen % should be acceptable. This should allow us to develop relatively fast feedback techniques which provide estimates within at most a few hours for a design. In this paper, we will concentrate on reducing the cost of power consumption, though other characteristics will typically be involved too in a practical context. In particular, a combination of area and power has to be targeted in real-time systems for a One of the most prominent eeects on the accuracy of early cost evaluation is the innuence of transformations on the initial high-level speciication entered by the system designer. The actual form of this spec-iication typically aaects the outcome of the architectural mapping stage very heavily. The estimate for a speciic measure can be factors or even orders of magnitude wrong compared to the optimally reach-able result when the speciication is not optimized rst. This problem applies when one uses accurate architecture mapping tools but also for high-level es-timators, and even for manual mapping as long as the initial control/data ow graph is kept. Still, it is required to obtain fast and early feedback on system design decisions in most industrial environments. Therefore, there is a need for design automation methods/tools which transform …