In-Situ Hydrogenated CMOS-Compatible WO3/Pd/PSG/WO3 ECRAM

To overcome the computational bottlenecks of large AI models, analog deep learning accelerators process information locally using special-purpose devices for matrix multiplication calculations and outer product updates in the analog domain. Among device candidates, Electrochemical Random-Access Memories (ECRAMs) modulate the resistance of a semiconductor channel through ionic exchange with a reservoir via an electrolyte. Ion intercalation is a non-volatile process with deterministic and reversible conductance potentiation and depression characteristics that makes ECRAMs most promising for neural network training with enhanced energy efficiency, non-volatility, and low latency. Our previous work [1,2] has demonstrated nanoscale protonic ECRAM devices enabling channel conductance modulation over a 20x range with nano-second operation. Our recent research is a close collaboration between MIT and IBM Research focused on developing a CMOS-compatible ECRAM process based on in-situ hydrogenated HxWO3 channel, PdHy gate reservoir, and phosphosilicate glass (PSG) electrolyte (Fig. 1a, b), which enables control of baseline hydrogen level and thereby device conductance as part of its fabrication.

References: [1] M. Onen, et al. Science, 2022, 377, pp. 539-543. [2] M. Onen, et al. Nano Letters, 2021. 21(14): p. 6111-6116.