Low-energy, high-accuracy convolutional network inference in 3D crosspoint (3DXP) arrays

In-memory computing (IMC) has emerged as a promising solution for artificial intelligence (AI) accelerators thanks to the reduced data movement and improved parallelism in the crosspoint memory array. A key issue of IMC is the excessive current of the memory elements causing energy inefficiency and computing inaccuracy due to IR drop. This work reports a hardware demonstration of IMC by a 3D crosspoint (3DXP) array of phase change memory (PCM). We experimentally demonstrate feature extraction, a typical layer of convolutional neural networks (CNNs) and simulate inference of a LeNet CNN for handwritten digits classification (MNIST database). Low energy is enabled by subthreshold operated 3DXP cells, while the high accuracy is supported by precise program-verify algorithms. The impact of read 1/f noise is discussed via measurements and simulations.