Artificial Neural Network Based Prediction Mechanism for Wireless Network on Chips Medium Access Control

As per Moore’s law, continuous improvement over silicon process technologies has made the integration of hundreds of cores on to a single chip possible. This has resulted in the paradigm shift towards multicore and many-core chips where, hundreds of cores can be integrated on the same die and interconnected using an on-chip packet-switched network called a Network-on-Chip (NoC). Various tasks running on different cores generate different rates of communication between pairs of cores. This lead to the increase in spatial and temporal variation in the workloads, which impact the long distance data communication over multi-hop wire line paths in conventional NoCs. Among different alternatives, due to the CMOS compatibility and energy-efficiency, low-latency wireless interconnects operating in the millimeter wave (mm-wave) band is nearer term solution to this multi-hop communication problem in traditional NoCs. This has led to the recent exploration of millimeter-wave (mm-wave) wireless technologies in wireless NoC architectures (WiNoC). In a WiNoC, the mm-wave wireless interconnect is realized by equipping some NoC switches with an wireless interface (WI) that contains an antenna and transceiver circuit tuned to operate in the mm-wave frequency. To enable collision free and energy-efficient communication among the WIs, the WIs is also equipped with a medium access control mechanism (MAC) unit. Due to the simplicity and low-overhead implementation, a token passing based MAC mechanism to enable Time Division Multiple Access (TDMA) has been adopted in many WiNoC architectures. However, such simple MAC mechanism is agnostic of the demand of the WIs. Based on the tasks mapped on a multicore system the demand through the WIs can vary both spatially and temporally. Hence, if the MAC is agnostic of such demand variation, energy is wasted when no flit is transferred through the wireless channel. To efficiently utilize the wireless channel, MAC mechanisms that can dynamically allocate token possession period of the WIs have been explored in recent time for WiNoCs. In the dynamic MAC mechanism, a history-based prediction is used to predict the bandwidth demand of the WIs to adjust the token possession period with respect to the traffic variation. However, such simple history based predictors are not accurate and limits the performance gain due to the dynamic MACs in a WiNoC. In this work, we investigate the design of an artificial neural network (ANN) based prediction methodology to accurately predict the bandwidth demand of each WI. Through system level simulation, we show that the dynamic MAC mechanisms enabled with the ANN based prediction mechanism can significantly improve the performance of a WiNoC in terms of peak bandwidth, packet energy and latency compared to the state-of-the-art dynamic MAC mechanisms