Analog Compute-in-Memory For AI Edge Inference

The information age has ushered in a wave of data at both the cloud and on the edge. Engineers have created new algorithms to identify higher-level information within this data, which often involves scanning with increasingly complex pattern recognition systems. Compute-in-Memory (CiM) is an architectural strategy that seeks to optimize these important applications by computing at the memory instead of inefficiently shuffling data to/from a CPU. This paper surveys a few of these techniques, then dives into analog CiM(ACiM) for edge AI, specifically looking at the Mythic M1076.