Hydrogen-Based Electrochemical Random-Access Memories (H-ECRAM) with Low Device-to-Device and Cycle-to-Cycle Variabilities

Programmable resistors with low variability are the key to achieve high training accuracy in hardware neural networks for effective artificial intelligence applications. Traditional candidates for programmable resistors, such as resistive random-access memory (ReRAM), often face challenges with large variability due to their stochastic nature of operation. Electrochemical random-access memories (ECRAMs) are promising three-terminal non-volatile programmable resistors with low variability, which is enabled by the deterministic and controllable dynamic doping of the channel material. However, the device-to-device variation has not been systematically studied. In addition, the effect of extended defects, such as grain boundaries in the channel, on the device-to-device variability is not clear. In this work, we systematically quantified the variability of CMOS-compatible hydrogen-based ECRAMs (H-ECRAMs) with crystalline and amorphous channels, and of various channel sizes. By programming multiple devices with identical parameters and examining their conductance modulation range, we observed very low variations in the low-conductance regime, including approximately 15% device-to-device and less than 3% cycle-to-cycle variation, with an endurance exceeding 10⁶ conductance updates. Furthermore, device-to-device variation showed no dependence on whether the channel was crystalline or amorphous, nor on channel sizes ranging from 10² μm² to 150² nm². These results demonstrate that ECRAMs meet the variability targets for programmable resistors and possess significant potential for downscaling, making them a promising candidate for use in hardware neural networks.