A configurable nonlinear operation unit for neural network accelerator

With the development of machine learning, neural network accelerators are widely used to speed up the calculation. Many of the accelerators are designed to be configurable so they can be widely used in different situations. Nonlinear operation is essential to a neural network with nonlinear fitting ability As there are a lot of optional nonlinear operations, a configurable nonlinear operation unit is necessary to a configurable neural network accelerator. This paper produced a configurable nonlinear operation unit used in neural network accelerators. It can realize different nonlinear operations by combining some basic operation units together. The connection mode among these basic units can be configured, so the nonlinear operation unit becomes very flexible. In addition, because of the re-use of these basic units, this circuit has an advantage of cell area and power consumption.