Harnessing Avalanche Criticality in Memristors for Versatile and Robust Computing

Avalanche criticality in nanoscale systems leverages switching events to trigger cascades across multiple scales, analogous to dynamics in biological neural systems. Criticality endows such systems with potentially large amplitude and frequency dynamic ranges. This allows them to process a wider range of inputs and adapt to varying computational tasks, which is highly beneficial for robust and versatile computation. However, nanoscale systems operating at criticality for practical computational tasks remain rarely demonstrated. Two challenges persist: (1) maintaining criticality while processing inputs spanning wide ranges, and (2) coordinating a large number of devices with substantial intrinsic device‐to‐device variation. For the first, we introduce a bias‐anchored dynamic input scheme together with a spike‐rate readout scheme to activate the large dynamic range of criticality in nanoscale memristors. This is verified by anomaly detection of bio‐signals, attaining accuracies above 90% under 2.7× amplitude and 5× frequency variations. For the second, we develop a device‐variation‐tolerant random source array that exploits the large dynamic range for token sampling in a large‐scale GPT‐2 language model, reducing device variability of output probability by 2.2× and enhancing token diversity by up to 2.1×.