Wireless NoC and Voltage Frequency Island Co-Design for Energy-Efficient Manycore Platforms
Kim, Ryan Gary
MetadataShow full item record
Multiple Voltage Frequency Island (VFI)-based designs present a scalable power management strategy for manycore chips. However, the overall communication backbone, which relies predominantly on Networks-on-Chip (NoCs), dictates the achievable performance. Emerging paradigms, such as the small-world wireless NoC (WiNoC), can be utilized to help improve the performance of manycore chips over traditional NoCs. The achievable energy savings in a VFI-enabled manycore system depends on the control mechanism for V/F tuning. A simple control module can statically tune the voltage/frequency (V/F) of each VFI to the average workload requirements of the application. This allows large energy savings while ensuring that the system is able to accomplish its task within the specified time-frame. However, most applications have time-varying workload requirements. VFI-based designs can take advantage of this time-varying nature through dynamic V/F allocation of each VFI. When dynamically adjusting the V/F of the VFIs, the state of the system must be partially or fully known; this involves the transfer of information from the cores to the V/F control module. Here, the WiNoC's higher bandwidth and low-latency communication is well-suited for efficient dynamic VFI (DVFI) control. Due to the variations within each VFI cluster, the selection of a single V/F that suits all cores within a VFI is a crucial and difficult problem. In this dissertation, we demonstrate that Machine Learning (ML) techniques can learn accurate DVFI control policies to improve the energy-efficiency of manycore systems. This DVFI control policy jointly predicts the V/F assignment for all VFIs by leveraging the structural relationships between them. In this dissertation, we demonstrate how the emerging WiNoC architecture is able to complement and enhance VFI-partitioned systems. By implementing a VFI-aware WiNoC, the penalties associated with inter-VFI communication are mitigated, DVFI control knowledge can be transmitted more efficiently, and the performance of the system can be significantly improved. Also, we have demonstrated control mechanisms that allow us, in conjunction with the VFI-aware WiNoC architecture, to achieve significant energy savings with practically no performance penalty. This opens up a new of class of co-design approaches that can make WiNoCs the communication technology of choice for future manycore platforms.