Energy Costs of Transporting Switch Control Bits for a Segmented Bus

When designing interconnection circuitry for Systems-on-Chip in deep sub-micron technology, it becomes clear [15] that the energy consumed by communication within a processing tile is gradually taking predominance over the energy consumed in the circuitry of the tile itself. Long wires are prevalent in most bus systems; therefore the segmented bus, which continuously minimizes the length of active wire, and therefore the capacitance that has to be driven, becomes attractive [20]. The segments of the bus are switched in and out of place by multiple switches, which themselves need to be controlled from the instruction subsystem of an instruction set processor. Whatever energy is gained in the data plane could well be lost in the control plane, unless measures are taken to minimize this. The control energy for a segmented bus has two main components: there is the energy required to distribute the control bits over the tile, and the energy required to fetch the control bits from the instruction memory. In this paper we concentrate on the first part: the energy required to distribute the switch control bits. We compare this energy with the energy consumed by the segmented bus itself, and also with the energy that would be consumed by the bus, were it not segmented. The switch control bits for a System-on-Chip need to be determined at design time by a suitable compiler. In the absence of such a tool, we have developed a method to determine the pattern of switch control bits and calculate the cost of driving the switches. We use a bus access schedule that is already power-optimized, and also a physical description of the bus structure and the modules attached. We do this for four alternative choices for the design of a digital broadcast receiver. The results indicate that the cost of driving the switches of the segmented bus do not offset the gain obtained from the segmentation, and that the prospects of driving the switches from the instruction processing subsytem are indeed promising, as long as the second part of the energy budget (fetching the control bits from the instruction memory) can be well optimized.