New ferroelectric device performs calculations within memory

In a new Nature Communications study, researchers have developed an in-memory ferroelectric differentiator capable of performing calculations directly in the memory without requiring a separate processor.

The proposed differentiator promises energy efficiency, especially for edge devices like smartphones, autonomous vehicles, and security cameras.

Traditional approaches to tasks like image processing and motion detection involve multi-step energy-intensive processes. This begins with recording data, which is transmitted to a memory unit, which further transmits the data to a microcontroller unit to perform differential operations.

Since differential operations are fundamental to several computing tasks, the researchers exploited the properties of ferroelectric material to create their device.

Tech Xplore spoke to co-authors Prof. Bobo Tian and Prof. Chungang Duan from East China Normal University. “We have been devoted to studying brain-inspired devices using ferroelectric materials for about ten years. Because of their non-volatile polarization, ferroelectric materials are commonly used for storage and emerging in-memory computing,” said Prof. Tian.

The von Neumann bottleneck

The foundation of modern computing lies in the von Neumann architecture. In such systems, the memory and processing units are separate, making them highly inefficient.

The data transfer between the processing units and memory causes latency and requires a lot of energy. This is known as the von Neumann bottleneck and is one of the more pressing issues with modern computing architecture.

Additionally, for certain tasks like image and video processing, the memory requirements are excessive as both current and previous frames are required for performing operations.

The researchers addressed these issues by leveraging the dynamic behavior of ferroelectric materials.

Ferroelectric capacitors

Ferroelectric materials have inherent polarization when an external electric field isn’t applied, which can be reversed upon application of the electric field.

Due to this dynamic behavior, ferroelectric materials can store and retain information in their polarization or alignment of dipoles. This is known as domain switching, with domain referring to a region of the material having a particular polarization.

“During ferroelectric domain switching, there comes measurable current signals, as ferroelectricity switching is essentially a change of the polarity of the dipole, which must generate electric current. This phenomenon is rare in other non-volatile materials where parameter change can only be detected by a following read operation,” explained Prof. Duan.

The researchers therefore decided to use ferroelectric capacitors as their differentiator devices. Capacitors inherently model change over time with the way charge is stored in them, making them an ideal candidate for differential operations.

Further, the way a capacitor stores and releases charge mimics memory. The capacitor remembers how much charge it holds until discharged, which translates to information storage as voltage levels across the capacitor.

This device is known as ferroelectric RAM or FeRAM. It is non-volatile like flash memory, meaning the device remembers the information even when the power is off in the form of polarization.

The future of edge devices

The researchers constructed a 40×40 passive crossbar array of 1,600 ferroelectric polymer capacitors. This means the device doesn’t have any other active components like transistors.

The capacitors can perform calculations directly, functioning as a RAM and CPU in a single device, eliminating the need for data transfer.

“Interestingly, the domain switching within a ferroelectric capacitor can generate macroscopically detectable currents in the circuit. When the ferroelectric domain orientation is encoded to store information, the domain switching gives in-situ differential information,” said Prof. Tian.

This means that the researchers use the current as a signal, directly indicating a change between successive inputs. Essentially, the device can identify differences between inputs without requiring additional calculations while also writing new data to memory.

The researchers demonstrated this capability through efficient motion detection in video processing and calculating the first and second-order derivatives.

The in-memory ferroelectric differentiator demonstrated energy efficiency, consuming about 0.24 femtojoules (fJ) per differential calculation when operated at 1 MHz (megahertz) frequency.

According to the researchers, their device is five to six orders of magnitude more efficient than present CPUs and GPUs, particularly the Intel 12900 and NVIDIA V100.

Due to their high efficiency, these devices could be excellent for edge computing applications, such as video and image processing, and biomedical devices for real-time processing of ECG/EEG data.

The technology’s scalability also appears promising.

“The absence of scaling constraints, thanks to silicon-compatible ferroelectric materials such as hafnia-based or aluminum nitride-based ferroelectric, allows mass production of ferroelectric arrays (＞1Gbit) capable of performing complex differential computations,” explained Prof. Duan.

The researchers explained that their long-term vision is to transition from data processing to physical law computing at the edge, where ferroelectric arrays natively resolve differential equations governing real-world phenomena.

© 2025 Science X Network