An In-Memory-Computing Charge-Domain Ternary CNN Classifier
Xiangxing Yang, Keren Zhu, Xiyuan Tang, Meizhi Wang, Mingtao Zhan, Nanshu Lu, Jaydeep P. Kulkarni, David Z. Pan, Yongpan Liu, Nan Sun
Abstract
The article presents a charge-domain computing ternary neural network (TNN) classifier with a complete four-layer neural network (NN) on a chip. The proposed ternary network provides 1.5-b resolution (0/+1/−1) for weights and activations, leading to 3.9× fewer operations (OPs) per inference than binary neural network (BNN) for the same Modified National Institute of Standards and Technology (MNIST) accuracy. The 1.5-b multiply-and-accumulate (MAC) is implemented by <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$V_{\text {CM}}$ </tex-math></inline-formula> -based capacitor switching scheme, which inherently benefits from the reduced signal swing on the capacitive digital-to-analog converter (CDAC). Also, the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$V_{\text {CM}}$ </tex-math></inline-formula> -based MAC introduces sparsity during training, resulting in a lower switching rate. The prototype is fabricated in a 40-nm LP CMOS process with an active area of 0.98 mm2, operates at 549 frames/s (FPS), and consumes 96 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mu \text{W}$ </tex-math></inline-formula> . With all OPs on the chip, it achieves 97.1% MNIST accuracy with 0.18 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mu \text{J}$ </tex-math></inline-formula> per classification, which is the smallest to our knowledge for comparable MNIST classification accuracy.