CVA NoC Project: Topologies
Network-on-Chip Topology Study
Our network-on-chip (NoC) topology study aims to determine the most power efficient network topology across a wide range of network parameters including network size, required bandwidth, traffic pattern, and relative energy cost of network components. In the experiments we compared the flattened butterflies (FBFly), Meshes, and Fat Trees topologies.
Figure 1 shows network power (in Watts) as a function of bisection bandwidth (bits/cycle) for networks with a fixed size of 64 nodes. The two FBFly networks consume less power over the entire range of bandwidths and their power increases more slowly with bandwidth than the other alternative topologies. At higher bandwidths, the 2D FBFly is preferred over the 1D.
Figure 1: Power vs. Network Bisection Bandwidth
Figure 2 shows how network power scales with the number of nodes in the network. In this experiment we held the bisection bandwidth of the network constant at 32 bits/cycle/node. Across a broad range of network sizes (8 nodes to 1024 nodes), the FBFly networks are still the most power efficient.
Figure 2: Power vs. Network Size
Figure 3 shows the effect of varying traffic patterns on the energy efficiency of network topologies. We simulated networks with 64 nodes and a bisection bandwidth of 2048 bits/cycle with seven different network traffic patterns. The power results shows that across all of these traffic patterns the FBFly networks are more efficient than the other topologies.
Figure 3: Traffic Comparison
Our study shows that regardless of bandwidth, network size, or traffic pattern, a network-on-chip using the flattened butterfly topology will always have lower power. Figure 4 is a diagram which shows the optimal dimensionality of a FBFly topology as a function of the network size (8-1024 nodes) and bandwidth (1-256 bits/cycle/node) under uniform random traffic. The figure shows that the optimal topology depends on both network size and bandwidth.
Figure 4: Power Optimal Topology Diagram
Contact
For further information, please contact Nan 'Ted' Jiang.