AxDNN: towards the cross-layer design of approximate DNNs

Thanks for the inborn error resistance of neural networks, approximate computing has become a promising and hardware friendly technique to improve the energy efficiency of DNNs. From the layer of algorithms, architectures, to circuits, there are many possibilities to implement approximate DNNs. However, the complicated interaction between major design concerns, e.g., power performance, and the lack of an efficient simulator cross multiple design layers have generated suboptimal solutions of approximate DNNs through the conventional design method. In this paper, we present a systematical framework towards the cross-layer design of approximation DNNs. By introducing hardware imperfection to the training phase, the accuracy of DNN models can be recovered by up to 5.32% when the most aggressive approximate multiplier has been used. Integrated with the techniques of activation pruning and voltage scaling, the energy efficiency of the approximate DNN accelerator can be improved by 52.5% on average. We also build a pre-RTL simulation environment where we can easily express accelerator architectures, try the combination of different approximate strategies, and evaluate the power consumption. Experiments demonstrate the pre-RTL simulation has achieved ~20X speed up compared with traditional RTL method when evaluating the same target. The convenient pre-RTL simulation helps us to quickly figure out the trade-off between accuracy and energy at the design stage for an approximate DNN accelerator.