Quantized Neural Network with Linearly Approximated Functions on Zynq FPGA
Miroslav Skrbek, Pavel Kubalík, Martin Kohlík, Jaroslav Borecký, Robert Hülle · 2024
This paper is focused on neural network implementation on FPGA. Linearly approximated functions combined with quantization are used to efficiently implement neural networks in hardware. Famous benchmarks were used for learning, evaluation, and hardware testing. Approximation-aware and quantization-aware learning were used to obtain weights for neurons in hardware. We implemented a neural network with an 8-bit architecture in VHDL and synthesized it to Zynq FPGA in Vivado. The resulting design running at 100MHz clock frequency was carefully tested against hardware-accurate models written in Wolfram Mathematica and C++. We present a decrease in FPGA resources and chip utilization compared to 16-bit architecture implementation.