Ultra-Low Power CGRA Architectures for General Purpose Acceleration at the Edge
Coarse Grained Reconfigurable Architectures (CGRAs) provide a better balance between reconfigurability, performance and area/power efficiency. Numerous CGRA architectures have been proposed targeting different application domains ranging from server class applications to biomedical/DSP applications. Major portion of the proposed architectures adhere to a common structure in which, homogeneous set of processing elements interconnected through a mesh network. This uniform resource allocation as well as unused flexibility results in a significant resource under-utilization which is more than 50% on average and this degrades the power/area efficiency.
We build up the hypothesis that a universal fully heterogeneous CGRA architecture can improve the utilization, thereby increasing the power/area efficiency. Such an architecture can be designed by restructuring the compute units and the network to be heterogeneous, supported by an efficient and distributed memory architecture. Heterogeneity will be supported by a compressed configuration memory system and a customized mapping algorithm. A fully heterogeneous architecture which is applicable for wide variety of applications will provide a significant improvement in power and area efficiency making it ideal for resource constrained environments.
