Publication date: 21st July 2025
Introduction
The emergence of so-called artificial intelligence (AI) applications is creating lot of buzz not only in the scientific community, but also in public media coverage. AI applications have made huge leaps in recent years and are capable of human-like text creation, conversation, and forms of reasoning. However, they come at a huge energy cost, with some sources estimating tens of millions of kilowatt hours of electricity use per day, and some companies already consuming more electricity than many countries in the world.
One of the main reasons for these exploding energy costs is the reliance of AI hardware on conventional von Neumann architecture with a separation of memory and compute elements. Despite the immense improvements in GPU, algorithms, and software efficiency, some current large AI models rely on over a trillion model parameters, which are stored in off-chip memory and need to be moved back and forth constantly between memory and compute. This movement of data makes up the main energy consumption.
Neuromorphic in-memory computing holds immense promise to obviate this need for data movement. The most fundamental building blocks of this approach are crossbar arrays, where each intersection of a number of mutually perpendicular lines holds a combined memory and compute element, often referred to as memristor. Such memristor crossbar arrays could carry out vector-matrix multiplications, the basic operation of AI models, directly in hardware. There are several approaches to achieve such memristor functionality. All have their advantages and challenges, and for reasons of industry-compatible materials and simple device fabrication, leading to low cost, this work is focussed on resistive switching.
Hybrid resistive switching
Among resistive switching devices, the majority of implementations, especially in industry-compatible materials such as silicon oxide or hafnium oxide, is based on a reversible soft dielectric breakdown, called filamentary switching [1]. In this approach, it is very difficult to control multi-level resistance states due to the ultra-fast and ultra-nonlinear switching characteristics of the process. Consequently, this approach suffers from challenges of uniformity. As an alternative, resistance states can be controlled by the voltage-controlled redistribution of ionic species inside a switching film. This provides better uniformity and resistance control, but often suffers from poor state retention and the switching can be too slow.
Over the past few years, we have combined these two approaches into a materials design concept which we term “hybrid resistive switching”, where we combine the strengths of both approaches to overcome each other’s challenges. We did this by depositing self-assembled nano-engineered thin films at industry-friendly 400 °C, which switch by forming “partial filaments” inside the thin film, but leave an area close to the bottom electrode unperturbed, and thus form an effective switching interface. This provides finely controlled multi-level resistance states with excellent overall characteristics.
Here, we present the implementation of our hybrid resistive switching materials design concept with two example materials. One is based on sodium bismuth titanate [2], a mixed ionic/electronic conductor of great interest for a wide range of applications, and the other is based on industry-friendly hafnium oxide [3]. Overall, we demonstrate switching speeds down to 20 ns, over 500 separate and stable resistance states across several orders of magnitude, retention measured up to 300 days, and spike-timing-dependent plasticity as demonstration of neuromorphic functionality.
While at the materials and device level, the results are very promising to take forward and make into crossbar arrays as the initial and fundamental building block of future fast and energy-efficient bespoke AI hardware.
We gratefully acknowledge funding by the ERC Advanced Grant EU-H2020-ERC-ADG #882929 grant “EROS”, the Royal Academy of Engineering Chair in Emerging Technologies CIET1819_24, and the EPSRC and U.S. NSF with grant number EP/T012218/1-ECCS-EPSRC.