Abstract

Approximate Computing (AxC) is promising in reducing circuit area, energy, and latency by slightly trading off accuracy for error resilient applications. However, AxC circuit quality control, i.e., to guarantee that the output error is within acceptable threshold, is still challenging. One fundamental reason is the lack of accurate error models of AxC applications. Most existing works perform static error analysis to get error statistics such as error mean and variance, which can be largely inaccurate and uninformative. In this work, we propose PreAxC, a novel error modelling and prediction flow for AxC designs. Instead of using simple statistics, we use error distribution for AxC circuit error analysis. We then propose graph neural network (GNN) based methods to predict the distribution based on the AxC programs, which can be effectively represented as data flow graphs (DFGs). We discuss two approaches: model-free and model-based, where the former directly predicts the error distribution histogram, and the later models the distribution using Gaussian Mixture Model (GMM) and predicts the GMM parameters. The experiment results demonstrate that our approaches can outperform existing error statistics and can successfully predict the error distribution, especially the model-free approach, even for completely unseen graphs (representing new AxC circuits) during training. The proposed PreAxC demonstrates the power of machine learning for circuit design, and is of great importance for future AxC design quality control.