Mixed-precision integer-arithmetic-only neural network-based equalizers for DML-based short-reach IM/DD systems

We propose mixed-precision integer-arithmetic-only NNs for low-complexity equalization in a 50-Gb/s 25-km DML-based PAM4 IM/DD link based on input neuron partitioning. The proposed scheme avoids floating-point weight operations and saves 26.3% memory at a minimum NN size compared with fixed-precision quantization counterpart without degrading BER.