International Conference on Solid State Devices and Materials

An automotive grade NOR-flash based in-memory computing architecture is proposed to execute similarity computation in analog domain. This analog computing chip can perform highly parallel analog in-memory searching (A-IMS) function using the specific encoding scheme to define valid data range and data proximity. The proposed two-block partition architecture can further double the analog search/data word length for high computing/ searching throughput. The one-shot channel hot electron (CHE) program operation enables fast analog data update. Experiments show that the A-IMS function has ultra-high immunity on data retention loss and read disturbance. Thanks to the analog computing algorithm, the chip can tolerate data VT distribution up to 3V in the 64-VT-level system and still keep high neural network image recognition accuracy up to 95%. The proposed NOR-flash-based A-IMS system is suitable for fast response memory centric AI applications.