FPGA routing congestion prediction combining DAGNN and GCN

Abstract

The issue of routing congestion is intractable due to the high complexity of FPGA (Field-Programmable Gate Array) design. Accurately predicting routing congestion in the early stages of the design process can shorten the overall design cycle. In the placement stage, this paper proposes an FPGA routing congestion prediction framework combining a deep adaptive graph neural network (DAGNN) and a graph convolutional network (GCN), the so-called DAGNN-GCN. DAGNN captures deep neighborhood topological information, focusing primarily on complex network features. Based on the hierarchical inclusion relationship between G-cells and cells, the extracted topological information is mapped geometrically onto the original geometric features and inflated to generate input feature vectors. These vectors are then fed into the GCN to train the routing congestion prediction of the model. Experimental results on the ISPD 2016 dataset demonstrate that the proposed method achieves Kendall, Spearman, and Pearson correlation coefficients of 0.57, 0.72, and 0.76, respectively, representing improvements of 16.33%, 5.88%, and 13.43% over the state-of-the-art LHNN method. Ablation studies further validate the efficiency of the proposed features in routing congestion prediction.