CNN Inference Using a Preprocessing Precision Controller and Approximate Multipliers With Various Precisions

This article proposes boosting the multiplication performance for convolutional neural network (CNN) inference using a precision prediction preprocessor which controls various precision approximate multipliers. Previously, utilizing approximate multipliers for CNN inference was proposed to enhance the power, speed, and area at a cost of a tolerable drop in the accuracy. Low precision approximate multipliers can achieve massive performance gains; however, utilizing them is not feasible due to the large accuracy loss they cause. To maximize the multiplication performance gains while minimizing the accuracy loss, this article proposes using a tiny two-class precision controller to utilize low and high precision approximate multipliers hybridly. The performance benefits for the proposed concept are presented for multi-core multi-precision architectures and single-core reconfigurable architectures. Additionally, a design for a merged reconfigurable approximate multiplier with two precisions is proposed for utilization in single-core architectures. For performance comparison, several segments-based approximate multipliers with different precisions were synthesized using CMOS 15nm technology. For accuracy evaluation, the concept was simulated on VGG19, Xception, and DenseNet201 using the ImageNetV2 dataset. This article will demonstrate that the proposed concept can achieve significant performance gains with a minimal accuracy loss when compared to designs that utilize exact multipliers or single-precision approximate multipliers.