345
Three-Dimensional Integration of Network-on-Chip
root routers in a 3D environment, the value of zero-load latency is lesser
than in 2D networks. But due to lesser number of inter-router links in BFT
networks, in actual traffic condition, simulation results show that both the
networks have almost similar latency profile under uniformly distributed
and localized traffic conditions.
The effect of traffic spatial localization on the average overall latency has also
been studied for all the networks under consideration (shown in Figures 11.13
through 11.15). The average overall latency of all the networks decreases with
increasing locality factor. As the locality factor increases, more traffic will go
to their local clusters. Hence, packets traverse lesser number of hops and will
create lesser contention in the network. It can be observed that the latency profile of both Mesh-2 and MoT networks in a 3D environment becomes closer to
their 2D counterparts with increasing locality factor. This is due to the fact that
the contention in all these networks becomes almost identical at highly localized traffic as they have single destination cores in their local clusters.
From the graphs shown in Figures 11.14 and 11.15, it can be observed that
the latency profile of the 3D 2 × 4 × 4 Mesh-1 network improves significantly
over the 2D 4 × 8 Mesh-1 network. Due to the rectangular structure of 2D
Mesh-1 network, packets traverse more hops in row-wise direction under
uniform distribution. Thus, the network suffers from more contention. In
3D topology, due to the square structure in vertical surface, the contention is
less. It has been observed in simulation that 3D Mesh-1 network has the best
latency profile under uniformly distributed and localized traffic condition.
11.3.3.4 Energy Consumption
Energy consumption in NoC is the summation of the energy consumed by
the routers and the communication links. Both these factors are network
topology dependent. Energy consumption of the proposed 3D MoT-based
network after applying clock gating in the FIFO under uniformly distributed
self-similar traffic is shown in Figure 11.16 for 200,000 cycles.
It can be observed that the network energy consumption increases linearly with the offered load but saturates as the offered load increases to
the throughput limit, similar to 2D NoC. Beyond saturation, no additional
packets can be injected successfully into the network and, consequently, no
additional energy is consumed. Simulation result shows that after gating the
write clock, the total energy consumption by all the FIFOs is about 35% of the
overall network energy consumption, whereas all the links consume almost
50% of it. The combined energy consumption by the routing logics arbiters
and control logic is about 15% of the total energy consumption.
Table 11.7 presents a comparison of average energy consumption at saturation for all the networks connecting 32 cores in 2D and 3D platforms. In
3D MoT network with dimension 2 × 2 × 4, the number of leaf routers having
connectivity 5 is 16 and that of stem routers (in vertical direction) having connectivity 3 is 8. In 4 × 4 2D MoT network, the number of leaf routers having