FPGA Routing Congestion Hot Spot Prediction Based On Graph Convolutional Network Transfer Learning

Physical design is an important step in the Field Programmable Gate Array(FPGA)design process.The high complexity of the design can often lead to routing congestion,which may result in non-malleability of the circuit.Recent studies have shown that predicting routing congestion is crucial for improving the quality of FPGA design.This paper proposes a routing congestion hotspot prediction method for FPGA design using a deep graph convolutional network model with transfer learning.First,we construct a graph convolutional network learning dataset.The FPGA benchmark is laid out using the UTPlace F layout tool,and the circuit layout is converted into a graph,with circuit node feature vectors extracted as node embeddings.The Vivado routing tool is used to perform global and detailed routing on the FPGA benchmark circuit layout,and the final routing congestion report is obtained.Based on the set routing congestion threshold,we analyze and generate the routing congestion hotspot labels for the generated circuit,and construct the graph convolutional network learning dataset with the node embeddings.Next,we formulate a routing congestion hotspot prediction model for FPGA design,which is reduced to a graph node classification problem.The optimal prediction model of the graph convolutional network is obtained through learning and training,achieving the routing congestion hotspot prediction for FPGA design.Finally,we establish a transfer learning method based on graph convolutional neural networks,including two methods: linear layer weight updating transfer learning and all weight updating transfer learning.We evaluate the performance of the transfer learning method under different numbers of training data samples.In this paper,the model based on a graph network and the transfer learning model based on a graph convolutional network is tested on the reference circuit of ISPD2016,and the experimental results show that the graph convolutional network model can effectively predict congestion hotspots,with an average accuracy of approximately87.5%,an average precision rate of approximately 95.53%,an average F1＿score of approximately 0.7705,and an average prediction time of 1/241 of the Vivado run time.The proposed transfer learning method can achieve the same performance as the non-transfer learning model with only 10% of the dataset,indicating that transfer learning can effectively address the problems of convergence and poor performance on small-scale datasets through knowledge transfer from the source domain.