Inference is typically much more power constrained, as it is often performed on a battery. For example in a mobile phone, or in an IoT sensor unit. This makes approximate computing for neural network accelerators an interesting field of research.