Low Temperature Modeling and Characterization of Analog ReRAM for Deep Neural Network Acceleration

Artificial intelligence is evolving raising significant challenges in terms of power consumption and processing speed. Deep Neural Networks (DNNs) on digital hardware require continuous data transfer between processor and memory, a problem known as the Von Neumann bottleneck. Bio-inspired computing systems, such as memristor devices arranged in a crossbar configuration, are alternative hardware architectures that can implement neural networks embedding weights in their conductance values. This configuration allows information to be stored and processed simultaneously, overcoming traditional bottlenecks. In parallel, the lack of cryogenic electronics introduces other challenges in the quantum computing field, where qubits need to be placed at cryogenic temperature to reduce thermal noise. This grows the interest in integrating electronics within the cryostat to reduce delays and improve system scalability, which is important for the development of efficient and scalable quantum computing systems. Traditional memories, like DRAMs, can operate in cryogenic environments as volatile memories, requiring periodic refresh operations. Emerging Non-Volatile Memories (eNVMs), such as Resistor-based memories, offer high scalability, low power consumption and refresh-free operation, making them promising candidates for cryogenic memory and analog in-memory computing. This work evaluates the viability of Resistive Random-Access Memories (ReRAM devices) to be used at cryogenic temperatures. The technology analyzed is a series of an Analog ReRAM, made by a bilayer stack of Conductive-Metal-Oxide(CMO)/HfOx between Titanium Nitride electrodes, and n-type MOSFET (1T1R structure). The transistor acts as a current compliance for the ReRAM that requires a fine control on the driven current. A cryogenic characterization is carried out both in quasi static and pulsed regime showing good performance down to 4 K. Moreover, the experimental data obtained are modeled by adapting the room temperature physical model to the cryogenic environment for a better understanding of the physics and transport mechanisms of the device.