Global power model

A DPU's output network consists of the RTL interconnect components connecting its output to the inputs of other DPUs. A data transfer between two DPUs is assumed to be center-to-center and rectilinear for the purpose of simplicity. [] shows that more than 95% of wires in an ASIC are routed in the Manhattan way, $i.e.$, without any detour from the shortest route. We therefore assume all data transfers are routed in the Manhattan way. A bus-based architecture shares the output networks among DPUs through busses. Since sharing output networks imposes a high performance and power penalty [46], we assume that output networks are not shared in the multiplexer-based architecture we use. Nevertheless, it is still possible to share wires within one output network. Fig. 7(a) shows one DPU sending data to four other DPUs through a fully dedicated output network. If the DPU sends all the data to all the dedicated data transfer wires, minimization of total power consumption in the output network will be the same as minimization of the total wire length in the output network. Therefore, a minimum spanning tree (MST) output network in the Steiner tree style [51] would have the least interconnect power consumption. However, if we take steps to reduce the SSA in the interconnect, minimal total length does not necessarily imply minimal total power consumption because it is hard to distinguish between dedicated and shared interconnects in the case of an MST. We introduce a trunk-branches style Steiner tree for the output network as shown in Figs. 7(b) and 7(c). The DPU sends all the data to a trunk which is either vertical or horizontal. All the other DPUs receive data from the shared trunk through perpendicular dedicated branches. Whether the trunk is vertical or horizontal depends on which one yields smaller power consumption. In Section V, we will compare three output network styles, fully dedicated, optimal total length shared and trunk-branches, in detail. We also assume all metal layers have the same capacitive parameters as the metal layer one. This assumption slightly underestimate interconnect power but obviates layer assignment and accelerate power estimation.

Figure 7: Different topologies of an output network: (a) fully dedicated, and trunk-branches output network with the trunk being (b) horizontal, and (c) vertical.