Computational Associative Memory Based on Monolithically Integrated Metal-Oxide Thin Film Transistors for Update-Frequent Search Applications
Zijian Zhao, Jorge Gómez, Huacheng Ye, Mohsen Imani, Xunzhao Yin, Shan Deng, Bryan Melanson, Jing Zhang, Xiao Gong, A. C. Abusleme Hoffman, Suman Datta, Kai Ni
Abstract
Associative memory based on ternary content addressable memory (TCAM) is typically used in a static inference mode where the update is occasional. However, for update-frequent associative search applications (e.g., clustering), existing TCAM designs have a fundamental gap between low-density, high write performance SRAM, and high-density, poor write performance nonvolatile memories. In this work, we demonstrate: i) monolithic 3D TCAM designs based on thin film transistors (TFT) that can simultaneously achieve high density and excellent write performance, thus bridging the performance gap among existing TCAM designs for update-frequent search applications; ii) the necessity of 6T TFT TCAM design to restore the desired independence among the TCAM words, which is destroyed in the 4T TFT TCAM design; iii) excellent write performance with logic-compatible write voltage (<1.5V), <20ns write latency, >10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sup> endurance; iv) up to 14x/35x improvement in speed/energy over GPU in executing the K-Means clustering algorithm, showing great promise of the TFT TCAM.