PCM-Based Analog Compute-In-Memory: Impact of Device Non-Idealities on Inference Accuracy
Xiaoyu Sun, Win-San Khwa, Y. S. Chen, C. H. Lee, H. Y. Lee, S. M. Yu, Rawan Naous, J.Y. Wu, Tao Chen, Xinyu Bao, Meng‐Fan Chang, C.H. Diaz, Hoilun Wong, Kerem Akarvardar
Abstract
The impact of phase change memory (PCM) device non-idealities on the deep neural network (DNN) inference accuracy is systematically investigated. Based on the experimental PCM data, statistical models of device non-idealities were extracted and incorporated into our PyTorch-based simulation framework for evaluations on the CIFAR-10 dataset. Our specific results include: 1) nonlinear <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${I}$ </tex-math></inline-formula> – <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${V}$ </tex-math></inline-formula> could incur a significant accuracy degradation, but it can be eliminated depending on how the input activations are encoded (e.g., no degradation with pulse-encoding schemes); 2) resistance variation and read noise induce a relatively mild accuracy degradation (< 1% with experimentally fit model), which can be further mitigated through variation-aware training (VAT); 3) maximizing accuracy over a given operating temperature range is attained through a “temperature-specific weight remapping” method developed in this work, accuracy variance of < 3% is demonstrated over a temperature range of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${T} \pm 15 ^{\circ }\text{C}$ </tex-math></inline-formula> ; and 4) resistance drift leads to a significant accuracy degradation over time and is the most challenging non-ideality to address by algorithmic means alone (drift coefficient < 0.015 is needed to achieve < 3% degradation in ten years). A “weight transfusion” (WT) method has been proposed to effectively recover the inference accuracy by incrementally activating additional pre-trained neurons over time. The main overhead is the additional area to store pre-trained weights beforehand, which is likely affordable given the high density of MLC PCM.