Review of Ferroelectric Devices for Smart Computing

Diagrams of (a) the challenges facing modern computers using the von Neumann architecture and (b) solutions for the “thermal wall”, “memory wall”, and von Neumann bottleneck based on ferroelectric devices . Credit: Intelligent Computing (2022). DOI: 10.34133/2022/9859508

Transistors or “microchips” are part of the reason why our paper-thin laptops can perform far more complicated tasks than their gigantic, clunky predecessors. To maximize computing capabilities, engineers try to make transistors as small as possible and pack billions of them into a single computer chip.

However, despite rapidly changing manufacturing techniques, traditional transistors are approaching their physical limit – these nanoscale devices cannot afford to shrink any further after a certain point – and this is hampering the development of computing capabilities. .

Yet, as data continues to pour in, the demand for computing capacity continues to increase. New devices, especially new storage and logic devices with higher speed and lower power consumption, are needed to unlock new computing capabilities while removing major barriers to existing computing systems.

Recently, a group of Chinese researchers pointed to ferroelectric devices as a promising solution and published a review article introducing emerging ferroelectric materials and devices for smart computing. The journal is published in Intelligent Computing.

Ferroelectric materials are quite versatile and widely used as special purpose memories in aerospace storage devices, among others. They have special polarization characteristics, a property similar to magnetism which can be retained even after the removal of the external electric field. But when the film thickness is reduced to less than 10nm, most conventional ferroelectric materials lose their polarization characteristics at 25°C and are therefore not suitable for the process of manufacturing integrated circuits (ICs) at scale. nanometer.

New ferroelectric materials with high scalability potential can solve these problems. “The discovery of the bias effect in high-κ materials, which are the commonly used gate oxide materials for nanoscale MOSFETs [metal-oxide-semiconductor field-effect transistors]is a breakthrough for the mass production of ferroelectric transistors,” the researchers emphasized.

They reviewed two important examples of polycrystalline Hf-based and amorphous oxide-based ferroelectric materials, and briefly described some recently reported new materials and devices. All are compatible with the CMOS (Complementary Metal Oxide Semiconductor) manufacturing process.

For advanced ferroelectric devices, the researchers categorized them into low-power logic devices, high-performance memory cells, and neuromorphic devices, and summarized them in detail. The abstracts covered the development of the devices and their abilities to break the “thermal wall”, the “memory wall” and the von Neumann bottleneck respectively.

Ferroelectric Negative Capacitor Field-Effect Transistors (NCFETs) as low-power logic devices are capable of breaking the “thermal wall”, which hampers the improvement of CPU core frequency due to power density increasing and the heating effect. “Reducing chip drive voltage is a potential method to break through the ‘heat wall’, and its feasibility is highly dependent on SS [subthreshold swing] of the transistor,” explain the researchers.

“Ferroelectric NCFETs, together with the voltage amplification effect, can overcome Boltzmann’s tyranny and achieve SS below 60 mV/dec. Thus, they are considered to have one of the most promising device architectures for very low power applications and can reactivate rapid development of the integrated circuit industry.”

Capacitor-based ferroelectric random-access memory (FeRAM) and ferroelectric field-effect transistor (FeFET)-based memory, classified as high-performance memory cells, exhibit excellent performance in dynamic RAM replacement. (DRAM) and embedded applications.

The ferroelectric capacitor, unlike the conventional DRAM capacitor, can store information via the Pr charge, which is non-volatile, and has a much higher charge density per area.

“Therefore, replacing the dielectric material of a flash device with HfO2-doped ferroelectrics or amorphous oxide ferroelectrics to achieve FeFET is an alternative method to further reduce the power or delay of these memories,” the researchers said. . This will help bridge a large performance or area gap between the logic device and the memory cell, overcoming the so-called “memory wall”.

Additionally, FeFETs can be used as neuromorphic devices to break the von Neumann bottleneck. Von Neumann bottleneck refers to delay and power problems caused by inefficient data transfer between the initially separate memory module and logic processor, to which neuromorphic computing – the imitation of the neural system for information processing – is a possible solution.

In a neuromorphic system, artificial neurons and synapses are the most important components, and FeFETs can implement both. For applications in neurons, FeFETs have been used as pulsed neural networks; For artificial synapse applications involving spiked neural networks (SNNs) and convolutional neural networks (CNNs), FeFETs are applicable due to their ability to simultaneously perform storage and processing functions.

Additionally, ferroelectric tunnel junctions (FTJs) have attracted particular attention for synaptic device applications due to their compact device structure, non-destructive read pattern, and high read/write access speeds.

In conclusion, the researchers indicated that if the trade-off between process compatibility and device performance can be achieved, NCFET, FeRAM or FeFET memory and ferroelectric synapse devices can be integrated in the same chip to build a multifunctional intelligent computer system.

“Based on advances in ferroelectric device processing technology, the integration of low-power logic, high-performance memories, and neuromorphic systems on a single chip appears to be achievable with continuous process improvement,” they said. they pointed out. “It will help realize the development of high-performance and high-efficiency intelligent computing systems in the future.”

Powered by Intelligent Computing