Uncertainty Quantification Based on Multilevel Conductance and Stochasticity of Heater Size Dependent C-doped Ge2Sb2Te5 PCM Chip

Uncertainty quantification (UQ) in deep learning plays a vital role in the process of decision making and optimization, and has practical applications in numerous fields of science and engineering. For the first time, we combine the multilevel characteristics of 4Mb phase change memory (PCM) with its inherent stochasticity to serve as weights in convolutional and stochastic layers simultaneously and hence achieve uncertainty quantification, where the intrinsic stochasticity of the device is utilized to efficiently capture the model uncertainty. The size of the heater electrode of PCM is varied to control the volume of phase change, where devices with larger 4nm heater generally exhibit less conductance drift, reduced state variation and better state linearity, making them more favorable for applications. This work improves the time and energy efficiencies in given uncertainty quantification benchmarks by 557 and $1.73\times 10^{6}$ times compared with GPU.