First demonstration of charge-domain content addressable memory based on ferroelectric capacitive memory for reliable and low-power one-shot learning

Non-volatile content addressable memories (NV-CAMs) accelerate the memory augmented neural networks (MANNs) by their in-memory-search capability and high parallelism, promising to realize brain-like efficient learning from a few examples or even one example with low energy consumption. However, most existing NV-CAMs operate in the current domain, which poses challenges in achieving reliable, low-power, and sensing-friendly Hamming distance (HD) computation essential for implementation of MANNs. To effectively address these key challenges, this work proposes transferring the computation to the charge domain using high-performance ferroelectric capacitive memory (FCM). For the first time, the demonstration of a charge-domain 2FCM CAM based on the inversion-type FCM with a large capacitance ratio is reported. By storing data as the device capacitance, this novel CAM structure is capable of directly outputting HD as stable multi-level voltages with a large margin and an ultra-high linearity, which significantly simplifies sensing processes and reduces peripheral costs. Moreover, the differential nature of the operation scheme further exhibits immunity to device variation, offering high accuracy in the computation of long data vectors. This work experimentally realizes parallel 16-bit HD computation using a fabricated 16×16 2FCM CAM array and reveals record performance at the array level: an ultra-low search energy consumption of 0.005 fJ/bit and a high accuracy of 97.5% on the classification of Omniglot dataset with one-shot learning. The proposed 2FCM CAM significantly outperforms the state-of-the-art current-domain CAMs, evidencing the superiority of charge-domain computation and showcasing tremendous potential for future in-memory-search applications.