Quantization Effects of Deep Neural Networks on a FPGA platform

In this paper, a quantization method for a FPGA platform is applied on three different deep neural networks (DNNs) for classification, detection and semantic segmentation tasks. The FPGA platform is based on a Xilinx Zynq UltraScale+ ZU5EV. The quantization process is done with the Vitis AI Quantizer tool of the AMD Vitis AI Environment. After a short description of the used DNNs, datasets and the technique behind the quantization process, the results are presented for different bit-widths of the quantized DNNs. The evaluation includes a post-training quantization of weights and activations in INT8 format with performance measurements in frames per second (FPS) on the target platform.