Network-on-Chip Topology Study
Our network-on-chip (NoC) topology study aims to determine the energy and area efficiency of network topologies across a range of network parameters including network bandwidth, traffic pattern, network frequency. In the experiments we compared the flattened butterflies (FBFly), the Mesh, the concentrated Mesh (CMesh), and Fat Trees (FTree) topologies. In this experiment we used a RTL based router model and spice based channel model to obtain the energy and area results. The router RTL were place and routed using a commercial 45 nm lower power library running at 200MHz. The channel model uses technology parameters from the same library.
Figure 1 shows network energy efficiency (in pJ/bit) as a function of network bandwidth for networks with a fixed size of 64 nodes running uniform random traffic. We change the network bandwidth by changing the channel width. The four data point for each topology corresponds to channel width of 16, 24, 48, 72 bits. For each channel width configuration, the network is running at 50% of saturation bandwidth. The flattened butterfly network provides better energy per bit over the entire range of bandwidths than the other alternative topologies.
Figure 1: Network energy per bit sent under uniform random traffic vs. network bandwidth
Figure 2 and 3 shows the effect of varying traffic patterns on the energy efficiency of network topologies. Each network configuration is running at 50% saturation throughput under the test traffic pattern. Flattened butterfly and mesh topologies uses dimension order routing while the fat tree uses randomized nearest common ancestor routing. In figure 2, using dimension order routing under transpose traffic, much of the network infrastructure is idle except for few heavily loaded channels. As a result the energy per bit of flattened butterfly and mesh topologies increases. Despite the reduced bandwidth under transpose, the flattened butterfly topology still has the lowest energy per bit within the range of channel width we tested.
Figure 2: Network energy per bit sent under transpose traffic vs. network bandwidth
In figure 3, nearest neighbor traffic heavily favors the mesh topology. Each node in the mesh has a dedicated channel to each of its immediate neighbors, this results in very high network bandwidth. For other more complicated topologies, nearest neighbor traffic under utilizes network resources such as the long channels of the flattened butterfly. As a result, these under utilized resources decreases the energy efficiency of these topologies when compared to the mesh.
Figure 3: Network energy per bit sent under nearest neighbor traffic vs. network bandwidth
Figure 4 shows the total network router area of each topology for 64-network with different channel width. The x-axis is the network saturation bandwidth under uniform random traffic. The flattened butterfly topology, despite using 10-port routers, gives the best area to uniform random bandwidth ratio among the topologies used in this study.
Figure 4: Total network router area vs. network saturation bandwidth
For further information, please contact Nan 'Ted' Jiang.