An ultrafast bipolar flash memory for self-activated in-memory computing

Abstract

In-memory computing could enhance computing energy efficiency by directly implementing multiply accumulate (MAC) operations in a crossbar memory array with low energy consumption (around femtojoules for a single operation). However, a crossbar memory array cannot execute nonlinear activation; moreover, activation processes are power-intensive (around milliwatts), limiting the overall efficiency of in-memory computing. Here we develop an ultrafast bipolar flash memory to execute self-activated MAC operations. Based on atomically sharp van der Waals heterostructures, the basic flash cell has an ultrafast n/p program speed in the range of 20-30 ns and an endurance of 8 × 106 cycles. Utilizing sign matching between the input voltage signal and the storage charge type, our bipolar flash can realize a rectified linear unit activation function during the MAC process with a power consumption for each operation of just 30 nW (or 5 fJ of energy). Using a convolutional neural network, we find that the self-activated MAC method has a simulated accuracy of 97.23%, tested on the Modified National Institute of Standards and Technology dataset, which is close to the conventional method where the MAC and activation operations are separated.