Multicast-Aware High-Performance Wireless Network-on-Chip Architectures

OrthoNoC: A Broadcast-Oriented Dual-Plane Wireless Network-on-Chip Architecture

“…Balancing the load alleviates this problem, but the throughput improvement still decreases after β 1 because the network cannot eject the broadcast flits fast enough. This is the case for most existing wired-wireless architectures, which integrate wireless interfaces at selected routers and use adaptive routing [10], [11], [29]: adaptive routing balances the load, but integration at the router level leaves the ejection links unchanged. We thus observe that increasing the bandwidth at the ejection links while balancing the load is necessary to maintain the throughput advantage beyond β 1 .…”

“…These mechanisms are similar to those of congestion-aware routing [29], but with two particularities: in ORTHONOC, they are implemented at the network interface and can be easily coordinated globally as all nodes have the exact same view of the events happening in the wireless plane.…”

“…In most wired-wireless architectures, these channels are used to implement orthogonal links between distant cores [7], [8], [29], [36]. Even if implemented with multiple channels, ORTHONOC considers a single broadband link shared by all cores instead.…”

“…These methods do not scale well with the number of participating nodes due to the high cost of introducing new channels or the increase of the token round trip time [37]. Better scalability is achieve with the protocol used in [29], where nodes request access by broadcasting short orthogonal request packets. However, this protocol is based on impulse radio techniques for which mmWave integrated solutions have not been explored.…”

“…In the literature, 65-nm transceivers with total circuit efficiencies of 7.3 pJ/bit [11], 1.95 pJ/bit [30], and 1.5 pJ/bit [31] can be found. Other works provide extensive discussions on the feasibility of more efficient transceivers implemented in 28 nm (1.3 pJ/bit, in [29]) or 22nm (≤1 pJ/bit, in [11], [32]). In light of these figures and applying empirical scaling projections [16], we obtain reasonable estimates at 45nm and 22nm.…”