14.1 A 22nm 104.5TOPS/W µ-NMC-Δ-IMC Heterogeneous STT-MRAM CIM Macro for Noise-Tolerant Bayesian Neural Networks
De-Qi You, Win-San Khwa, Bo Zhang, Fang‐Yi Chen, Andrew Lee, Yu-Cheng Hung, Yiming Li, Yuhui Wang, Chung‐Chuan Lo, Ren-Shuo Liu, Kea‐Tiong Tang, Chih-Cheng Hsieh, Yu-Der Chih, Tsung-Yung Jonathan Chang, Meng‐Fan Chang
Abstract
Compute-in-memory (CIM) macros [1]–[5] for convolutional neural networks (CNNs) [6]–[7] and vision transformers (ViTs) [8] enable high-performance computing for energy-efficient edge-AI devices. Image recognition applications are constrained by inference accuracy degradation or <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{misjudgments}$</tex> due to environmental noise. Bayesian neural networks (BNNs) [9], which represent weights using its mean (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex>-weight) and difference from mean tF <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\Delta$</tex>-weight) outperform CNN and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{ViT}$</tex> models in terms of noise tolerance, making them promising candidates for edge-AI devices dealing with noisy inputs. Using digital circuits to implement BNN [10]–[12] imposes a tradeoff between inference accuracy and performance, power and area (PPA).