Iterative Multiplication Unit for Inference and Train Neural Networks

DNN deployment on the embedded system requires generality in embedded hardware to produce an optimal performance on DNN models quantized for different precision. The embedded system should be able to fine-tune the model at the edge to mitigate environmental effects. Generally, the training of DNNs requires higher precision compared to inference. As a result, few embedded devices support DNN training. An accelerator designed for DNN training underperforms in inference tasks. This paper presents an iterative MAC unit, where higher precision arithmetic calculation can be performed iteratively using a lower precision MAC unit. The iterative MAC unit can be used to implement a DNN inference model (with a single iteration) and can also be used to train the DNN model (with multiple iterations), which requires higher precision calculations. The proposed MAC unit can provide good performance and flexibility. Error resilient nature of the neural network is explored to reduce the number of iterations required in training. The effectiveness of the proposed iterative MAC is evaluated by simulating the LeNet-300-100 model.