Best-in-Class Density Single-Ended ePCM Memory Array for Weight Storage in Edge-AI Applications

The enablement of a high-density memory array in a System on Chip (SoC) is a great opportunity to overcome the Von-Neuman bottleneck in computation, especially for applications at the edge. In this work, we will show that a dense array based on 0.019μm² Phase-Change-Memory (PCM) cell, embedded in 18nm FD-SOI CMOS process, typically operated in differential mode, can be operated in single-ended mode for effectively storing Neural Network (NN) weights on a wide range of mission profiles. Differently from the previously adopted differential approach, the single-ended is less resilient to PCM reliability mechanisms, so proper algorithm and design optimizations have been carried out to provide a reliable behavior.