MoS2 Flash Memory Arrays with Sb Contact for Highly Efficient and Low-Latency Analog In-Memory Searches

With the rapid growth of AI and edge devices, the transfer of data between memory and computing units has emerged as a significant bottleneck in traditional digital hardware. Content addressable memories (CAMs) offer a promising solution by processing information directly within memory, but existing implementations using SRAM and, more recently, those using emerging nonvolatile memories are constrained by the performance of silicon transistors. In this work, we introduce an ultra-compact analog CAM that utilizes atomic thin 2D MoS₂ flash memories with semimetal antimony (Sb) contacts. Our device achieves the highest readout current (60μA/μm) and large ON/OFF ratios (>109) reported in 2D flash memories, which results in record low energy consumption and latency (below 0.1fJ/10 ps per search per cell) during analog in-memory search operations in our physical 8x16 analog CAM array with 256 MoS2 flash memory devices. Furthermore, we showcased analog Hamming distance computing for k-nearest neighbor (KNN) classification using our 2D analog CAM hardware, showing high accuracy, energy efficiency, and low latency for machine learning applications. This work experimentally demonstrates the potential of 2D materials in overcoming current hardware limitations, paving the way for more efficient and scalable computing solutions in intelligence edge devices.