In-Memory Computing array using 40nm multibit SONOS achieving 100 TOPS/W energy efficiency for Deep Neural Network Edge Inference Accelerators

摘要

40nm SONOS (Si-Oxide-Nitride-Oxide-Si) based non-volatile memory (NVM) cell has been evaluated for analog memory to perform in-memory neuromorphic computing. Process flow and smart-write algorithms were developed to tune key parameters like sigma, retention and noise performance for this application. Their optimizations to meet the product reliability requirements are also discussed. The performance and power of SONOS in a dense array is evaluated, and the sigma and retention data obtained are discussed.