COPS: An Efficient and Reliability-Enhanced Programming Scheme for Analog RRAM and On-Chip Implementation of Denoising Diffusion Probabilistic Model
Zhixing Jiang, Yue Xi, Jianshi Tang, Yuyao Lu, Ruihua Yu, Ruofei Hu, Qi Hu, Bin Gao, He Qian, Huaqiang Wu
Abstract
This work develops an efficient and reliability-enhanced programming scheme for analog RRAM, named Conductance and Operation-dependent Programming Scheme (COPS), which maximizes the probability of RESET as the final operation with small last conductance change (ΔG <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">last</inf> ) and further incorporates verify-delay-verify (VDV) to effectively suppress the post-programming relaxation effect of RRAM. Compared to the standard incremental step pulse programming (ISPP) scheme, COPS reduces the relative deviation (RD) of conductance drift by 38.6% at the same programming speed. Meanwhile, it reduces the programming time by up to 55.8% compared to the repeated programming (RP) with the same RD. The effectiveness of COPS is verified on different RRAM materials stacks and further validated by the on-chip classifications of handwritten digits and news articles, achieving 3~4-bit software-equivalent accuracy over 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">4</sup> s. For the first time, a more drift-sensitive task, denoising diffusion probabilistic model (DDPM), is further implemented on-chip, yielding results comparable to 4-bit software accuracy.