ULBNN: Ultra-lightweight BNN Accelerator for Out-Of-Order Execution in Autonomous Driving

The development of BNN (Binary Neural Networks) significantly reduces the inference computing complexity and storage overhead of the model, which has enormous promise for vehicle platforms with limited resources. In-vehicle platforms focus more on the power consumption, delay, and storage overhead of the model, while the accuracy requirements only need to reach a predetermined threshold. This paper introduces two optimization strategies based on threshold pruning technology in BNN. Firstly, we propose a regional differentiation training method. Secondly, we recommend using an out-of-order execution scheduling algorithm. We implement this design on the FPGA platform and use traffic signal identification datasets (such as GT-SRB, CTS, and BTS) to test the performance of the model. The research results show that the model can reduce operations by 37.9% without sacrificing accuracy, thereby achieving 1.85× the inference acceleration than the unoptimized model. Additionally, the model exhibits a 40% improvement in energy efficiency on the same data set when compared to the most sophisticated BNN model. We also tried universality testing on CIFAR-10 datasets, and the model performance was comparable. For hardware BNN accelerators, experimental results show that the performance of this work is better than other FPGA-based accelerators, and the energy efficiency ratio of the BNN inference reaches 3.21× to 4.54×.