Lifelong Learning with Monolithic 3D Ferroelectric Ternary Content-Addressable Memory
Soumya Dutta, Abhishek Khanna, Huacheng Ye, Mohammad Mehdi Sharifi, Arman Kazemi, Matthew San Jose, Khandker Akif Aabrar, J.G. Mir, M. Niemer, Xiaobo Sharon Hu, Suman Datta
Abstract
Lifelong learning at the edge requires on-the-fly learning from scarce data in one or few shots. Here, we present the array-level demonstration of few-shot learning using a first time fabricated monolithic 3D ternary content addressable memory (M3D-TCAM) using back-end-of-line (BEOL) ferroelectric FETs (FeFETs). The fabricated two-tier structure consists of two 10×10 sub-arrays in each tier and allows massively parallel search operation up to 20-bit long search vectors. We experimentally demonstrate: (a) record low write voltage of ± 1.6V with 20ns write latency for BEOL FeFETs in M3D-TCAM arrays, (b) in situ computation of Hamming distance (nearest neighbor match) between a 20-bit search vector and ten different stored vectors, (c) disturb-free write operation, and (d) high write endurance exceeding 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sup> cycles. We experimentally demonstrate a 3-way 3-shot learning with 20-bit feature vectors using Omniglot dataset and achieve an inference accuracy of 70% comparable to GPU accuracy of 72%. System-level benchmarking performed on a 64×512 M3D TCAM with 8 vertically stacked sub-arrays exhibit a 3.5x, 3.7x, 3.5x and 12x improvement in area, search energy, write energy and write latency, respectively, over 2D TCAM.