Publication date: 15th December 2025
Artificial intelligence is pushing the limits of digital computing to such an extent that, if current trends continue, global energy consumption from computation alone could surpass all other forms of energy use within the next two decades. Electrochemical random-access memory (ECRAM) presents a promising approach to storing and processing information in analog form, increasing efficiency and sustainability with in-memory computing [1-3].
In this discussion, I will introduce a novel approach to ECRAM, where devices are self-heated to overcome the kinetic barriers associated with electrochemical reactions. We refer to these devices as electro-thermo-chemical random-access memory (ETCRAM) [4]. The key innovation is an electrothermal gate that simultaneously distributes heat and facilitates oxygen vacancy reactions, enabling large reversible modulations of composition with a tunable analog resistance of up to one-billion times (
We demonstrate that programming an ETCRAM cell based on TaOx achieves current-voltage linearity across the entire operational range, a feat not possible with other memory technologies. This linearity enables a wide variety of signal processing tasks inaccessible to other devices and is attributed to a unique set of physical properties, including a non-electrostatic, volumetric programming mechanism and distinct read and write paths. Importantly, the self-heating mechanism significantly reduces noise, resulting in precision that is
We showcase a variety of signal processing tasks, including dynamic-gain amplification (up to 1V) and voltage-encoded matrix-vector multiplication (MVM), both demonstrating low harmonic distortion. Furthermore, we present the integration of a 1,024-element ETCRAM array with CMOS technology. Due to its unique linearity and precision at high resistance, simulations indicate that MVM efficiency could approach >1,000 TOPS/W.
