All-digital time-domain CNN engine for energy efficient edge computing

Machine Learning is finding applications in a wide variety of areas ranging from autonomous cars to genomics. Machine learning tasks such as image classification, speech recognition and object detection are being used in most of the modern computing systems. In particular, Convolutional Neural Networks (CNNs, class of artificial neural networks) are extensively used for many such ML applications, due to their state of the art classification accuracy at a much lesser complexity compared to their fully connected network counterpart. However, the CNN inference process requires intensive compute and memory resources making it challenging to implement in energy constrained edge devices. The major operation of a CNN is the Multiplication and Accumulate (MAC) operation. These operations are traditionally performed by digital adders and multipliers, which dissipates large amount of power. In this 2-phase work, an energy efficient time-domain approach is used to perform the MAC operation using the concept of Memory Delay Line (MDL). Phase I of this work implements LeNet-5 CNN to classify MNIST dataset (handwritten digits) and is demonstrated on a commercial 40nm CMOS Test-chip. Phase II of this work aims to scale-up this work for multi-bit weights and implements AlexNet CNN to classify 1000-class ImageNet dataset images