Efficient Design of Prefix Adders Based on Relational Graph Convolutional Network

Due to the complex topological structure of prefix adders, a set of design solutions typically contains a large number of non-dominated solutions. Commonly, to identify the dominated solutions, the conventional design process involves performance evaluation of the design outcomes, a procedure that consumes a significant amount of time. To address this, this paper proposes a new prefix adder design process based on graph neural networks, aimed at increasing the proportion of dominated solutions in the design outcomes. By utilizing Relational Graph Convolutional Networks (RGCN) to train existing prefix trees and create predictive models, we modify the feature matrices of known prefix graphs to generate new prefix trees that potentially offer improved performance. Experimental results indicate that our method has increased the proportion of dominated solutions in the generated solutions to 0.7%, while reducing the number of evaluations for adder performance by 47.8%. This approach adheres to the rigorous standards expected of high-level research in the fields of neural networks and circuit design.