FPGA Based CNN Accelerator Using Extended Exact Booth Multiplier

AI deployment on edge device is an emerging research field and is crucial in terms of performance, real time processing, latency and energy efficiency. But such applications present the challenges of high model size and computational complexities of the architectures. This work addresses these challenges by delving into concepts of approximate computing and custom bit-width quantization schemes. The proposed re-encoding scheme is based on the traditional Booth encoding that can reduce the model size with minimal loss in accuracy. The advantage of the re-encoding scheme is demonstrated by computing a Convolutional Neural Network (CNN) using an extended exact multiplier which is designed using the re-encoding scheme. The proposed re-encoding has reduced the model size of CNN by 71.15% with minimal loss in classification accuracy. Additionally, the proposed multiplier reduces the error rate when combined with the re-encoding scheme. This multiplier when integrated with neural network, achieves the best trade-off among accuracy, energy efficiency and power consumption making it suitable for deployment on edge devices.