3D-NAND based Filtering Cube with High Resolution 2D Query and Tunable Feature Length for Computational SSD
Po-Hao Tseng, Shao-Yu Fang, Yu‐Hsuan Lin, Feng-Ming Lee, Jhe-Yi Liao, Yu‐Yu Lin, Ming-Hsiu Lee, Kuang-Yeu Hsieh, Keh-Chung Wang, Chih-Yuan Lu · 2024
A novel 3D filtering cube is proposed in 3D-NAND flash for computational SSD to execute 3D in-memory-search (3D IMS) function at ultra-high search bandwidth. The 3D filtering cube compares 2D data inputs from word lines (WLs) and string-select lines (SSLs) with the database stored in the 3D NAND cubes. Multi-block operation using string select line (SSL) modulation provides tunable search word length for various dataset applications. Two encoding algorithms were developed for approximate and exact-range computing modes, which provide high filtration efficiency up to 99% from SSD to CPU (or GPU). The cube with normal encoding scheme features ultra-high search bandwidth up to ~3.5Tb/sec. A novel reverse encoding scheme was developed for enhancing output resolution with current-mode page buffer design. Reliability concerns including retention, read disturbance, page buffer (PB) design, and data pattern effects were evaluated. The concepts were successfully demonstrated with 5.5-bit word resolution in our 96-layer 3D NAND chip. The 3D filtering cube can facilitate minimal data movement and be a key component in computational SSD and memory-centric AI computing.