World Record ReRAM and More

Photo of Ron Neale, Renowned Phase-Change Memory ExpertRon Neale joins us to look at a new World Record for Oxide-ReRAMs.  Here he explores the possibility that, rather than an analogue-like device with a continuum of conduction states, the resolution limit of conductance stems from discrete building blocks or nano-filaments that define the changes between those states.  A dichotomy of views on oxide ReRAM operation is part of the mix.

For Oxide ReRAM watchers, one of the highlights of the first half of 2023 was a new World Record, published in the Journal Nature, that was claimed as the result of a joint effort by a team from the University of Massachusetts, Amherst, the Massachusetts Institute of Technology, Cambridge University, the University of Southern California, Los Angeles, and TetraMem Inc.

This team was able to put a full 2,048 conductance levels into a single ReRAM cell with the stable levels clearly separated by a conductance of 2 micro-Siemens (µS).  This feat perhaps received less media attention than it deserved.

As well as memory, that work had as its target neuromorphic AI applications and in-memory computation for ReRAMs.  Although the research effort was aimed at solving the problems associated with those applications, it also has provided some important new insight into the operation of HfO based ReRAMs.

Behind that World Record is a technique, given the formal name of “Denoising,” which solves a fundamental problem.  Denoising is performed after a conventional SET operation attempts to SET the bit to a target conductance level, as would be used for a two-level memory.

Denoising then involves what might be described as a clean-up operation to remove the causes of random telegraph noise (RTN).  In the past, it is RTN which has prevented large numbers of different conductance levels to be stored stably in a single device.

The Denoising clean-up uses a sequence of small multi-level pulses applied to the ReRAM while in its target level state.

One of the interesting insights from this latest work is the shape of the SET state area both before and after the Denoising operation.  The image below shows C-AFM (Conductive Atomic Force Microscopy) conductance maps across the surface of a ReRAM in its SET state, for two examples of a device before and after Denoising.

Four highly-colored images produced by C-AFM of a bit cell after programming. Top View. The conductive paths are red, the nonconductive areas are blue, and other colors represent states between these two. The top two are before and after de-noising for one bit, and teh bottom two for another. In both cases a somewhat circular pattern of red points exists with verious extra red points off to the side, or blue points within the red area before de-noising. After de-noising these are gone.

The dark blue represents insulating areas, red shows areas of high conductance, and other shades represents conductance between those two extremes.

In (a) initially there are two areas of high conductance, one being a small distinctly separate outlier in the dotted circle, which is removed (b) after Denoising.  While in (c) regions where robust high conductance bridges have not been made, those bridges are completed by the Denoising process (d).

An important feature is that the outline shape of the whole of the conducting area for the SET state is not represented by a neat cylinder, which expands symmetrically in radius with each increase in SET current.  Instead, the periphery appears to grow in a random fashion, with each new robust connection being made via the addition of what might be considered as an individual building block, possibly nano-filaments, at the periphery, each located where the next-easiest path is available.  Overall, the initial electronic threshold switching event and subsequent avalanche of electrons and oxygen ions combine to localise the area where the changes can occur.

The World Record was produced by a ReRAM with 2,048 SET states, whose conductances ranged from 50µS to 4,144µS.  If we assume a resolution of 2µS, the lowest conductance (50µS) would have 25 of the incremental building blocks, while the highest conductance (4,144µS) would have, 25 + 2,047 = 2,072.

In the calculation of the previous paragraph, we must consider the possibility that the 50µS state uses zero building blocks, and represents the conductance of the device in its low conductance state.  This material that will always be electrically in parallel with the SET states, is the dark blue material in the C-AFM image.

Does a resolution in conductance of 2µS represent a fundamental limit for ReRAM building blocks if they are separate entities?  “Yes and no,” or even “possibly,” might be the quick answer.

If we consider the ReRAM to be an analogue device, then it should be possible to find a whole new set of states between each of the states of the World Record example of the previous paragraph by say 1µS, although instability might prove to be a problem.

If, on the other hand, the conductive state is made up of discrete building blocks, or some minimum size nano-filament percolation paths, rather than an analogue continuum, is it possible to get some estimate of these building blocks’ dimensions?  Some clues might be found by exploring the shape of the periphery of the conducting region in the C-AFM, and the dimensions of the outlier in image (a).

The following offers an alternative.  If the roughly-elliptical-shaped conducting area in image (b) is considered to have major and minor dimensions of 8.33nm and 6.6nm, then the area would be 174 x 10-18 m².  If we then assume that area is capable of accommodating 2,072 discrete states, dividing by 2,072 results in a radius for each building block or nano-filament of 1.63 Angstrom.

While that result might just be good fortune, that is close the length of the Si-O bond.  This, in turn, suggests that each of the 2,048 SET states is achieved by the addition of a conducting string or individual percolation path consisting of vacancies made from what would have been a series of Oxygen-Si-Oxygen bonds, each with a diameter of 3.26 Angstroms.  This would then suggest the 2µS is most likely close to the resolution limit for the SET state.

One Large or Many Nano-Filaments?

For a filament-based model, the picture that emerges is illustrated in the illustration below.  It shows a second-generation device structure with an active volume of thickness 5nm or less, consisting of four films: the bottom electrode (green), the oxide (blue), a thin interface film (pale yellow), and the upper electrode (green).  This is the structure used by the World Record team.  First-generation ReRAMs used substantially thicker films.

A side view (top) and top view (bottom) of a single bit cell using the thoughts expressed in this blog post. The side view represents that 5nm-thick film between top and bottom electrodes as a fade from blue in the lower part to yellow in the top. Conductive columns are drawn in as red, from a bottom red part common to all columns to a top, which goes into a yellow-orange area. A couple of broken paths are illustrated as grey on top and blue-white at the bottom. An inset photo from the first figure compares its C-AFM to the drawn top view.

The cylinders in this illustration are my speculative representation or a possible interpretation of the 2µS increments in conductance.  The tiny red line with arrow head though each cylinder represents a single conducting percolation path through the vacancies in this model.

The cylinders that are grey on top and blue-white below represent incomplete or blocked nano-filaments or conductance paths.

The overhead projection of the nano-filaments in the figure serves as a link from the filaments to the C-AFM scan shared by the World Record team, shown on the inset.

The irregularly shaped outline of all the conductance increments would seem to contradict the concept that the conducting element consists of a single large-area filament.  The single large-filament model lacks an explanation of why in the C-AFMs the undulating edges of the conducting SET area is defined by what appear to be the outline of a series of small filaments.  In a single large filament model, incremental changes in conductance would result from the movement or removal of vacancies from somewhere in the volume of that single large filament, not only at the edges.

One important point of Figure 2 is that whether the conducting path of the SET state is a collection of nano-filaments or a single large conducting filament, its length is very small and is closely associated with the interface film (yellow) between the upper electrode and oxide.

Bulk conductance for a single large-area filament in the SET state can be viewed as a collection of individual conducting percolation paths formed by vacancies, electrically in parallel.  They approximate to nano-filaments.

A given step change in bulk conductance could be represented either by a single conducting path interruption in different percolation path locations throughout the volume of material, or by multiple interruptions of a single path.

This illustration uses two colours to represent two conditions of the interface film.  Pale yellow represents the condition after Forming which might be the result of changes that extend across the whole area of the device.  The yellow-orange, represents changes that have occurred when the device is in its SET state.  This volume is “Oxygen or Vacancy rich,” and conducts, behaving as an electrode when the cell is in both its SET and RESET states.  As we will see in the following section, some researchers support the view that it stores oxygen while other researchers believe that it is a vacancy reservoir.

The idea of building blocks is not new.  Those familiar with the field will be aware that as far back as 2015 a paper published in Nanoscale by Mark Buckwell et al., The Royal Society of Chemistry did provide evidence of multiple filament formation.  Even earlier, a 2012 paper Adnan Mehonic et al. suggested that the conducting state could consist of column-like cylinders formed from strings of conducting vacancies on the surface of oxide columns which bridge the inter-electrode gap.

An important contribution that supports the possibility of the existence of very small building blocks is the observation by the World Record team, that an individual increment in conductance could be inhibited by even a single trapped electron.

Another aspect of the second-generation ReRAMs is the thickness reduction of the column of permanently-conducting material, which was a feature of earlier thick film oxide devices.  In the illustration this is represented by the solid red layer below the filaments.  The move to active films with thickness equal to or less than 5nm means this permanently-conducting material has gone, or reduced to what in effect are a few nucleating points on the surface of the lower electrode.  In effect, it is the lower electrode.  In older thick-film designs the solid red permanently-conducting layer was thought to have filled the bulk of the space between the electrodes.

ReRAMs: A Dichotomy of Viewpoints.

While the ReRAMs based on the oxides of Ta, Ti, Hf, Si, etc. tend to get lumped together with a single description of the way they operate, there are two mechanisms used to account for the two different or multiple memory states.

The terms used to describe the different types of ReRAMs are “Oxygen-Ion Drift,” and “All-Vacancy,” names which, in brief, describe the key feature of operation.

While the observation of two terminal electrical characteristics might make either device appear to be an ideal form of memory, success will depend on an agreed-upon scientific explanation of the inner workings, or at least an understanding of which family they belong to.

The two oxide ReRAM types both have an agreed starting point, which is the injection of electrons into defects formed from distorted O-(M)-O or O-Si-O bonds.  This weakens those bonds in such a way that makes it possible to use an electric field alone to break them.

The result of that bond breakage is an avalanche of oxygen ions (anions) and electrons which move towards the positive electrode, leaving in their wake NV conducting material, where an increase in oxygen vacancy density provides the means of conduction.

That would appear to be an agreed position.  From then forward, and for normal memory operation, things change.

Four sketches to represent the low (left) and high (right) resistance states of two kinds of ReRAM: Oxygen-Ion drift (top two) and All-vacancy (bottom). The figure uses the same layers as the prior diagram: Green at the top and bottom for electrodes, and a fading ara from blue (lower) to yellow (upper) for the active film. Ions are drawn as circles to represent the ion movement in the medium.

The figure above in a simple way illustrates those changes.  At the top we have the Oxygen-Ion Drift type made of some form of silicon oxide, and on the bottom is the All-Vacancy type, made of HfO.  For both types, the left shows the Low-Resistance State (LRS) and the right shows the High-Resistance State (HRS).

As illustrated in the top left of this figure, in the Oxygen-Ion Drift type, after Forming the oxygen ions (blue circles) have moved into the region of the interface film (yellow, labelled “Getter”), leaving in their wake a conducting film of vacancies (white circles).  The device is in its low resistance (LRS) SET state.  The area where these oxygen ions have settled is something that was referred to early in this post as an Oxygen rich zone.  It has been coloured yellow-orange.

From then forward, and for normal operation, each time the applied voltage is reversed, the backwards and forward movement of the oxygen defines the high and low resistance states.

For the All-Vacancy mechanism of operation at the bottom of the illustration, the Forming process appears to be similar except that the oxygen disappears from the scene or is no longer involved in the normal operation of the device.

Two explanations are offered for the oxygen’s disappearing act: one is that it moves, possibly along grain boundaries of a porous electrode structure, to the outside environment, the second is that it moves into permanent strong bonding sites in the upper electrode and cannot be returned by the subsequent applied voltage.  A paper by Prof. Kenyon certainly provides evidence of the ability of oxygen to move into the electrode along grain boundaries.

From then forward, during the memory device’s normal operation, the movement or removal of vacancies defines the two or more resistance states.

The interface layer plays different roles depending on the type of cell.  In the Oxygen-Ion Drift type, it serves as a temporary reservoir of oxygen for the low resistance state of the memory.

In this device the interface film is an oxygen-attracting metal.  For example, Ti might be the choice when the main electrode is Tantalum (Ta).  That interface film is often described as a “Getter,” reinforcing its role in the operation of the device as a material with an affinity for oxygen.

(The word “Getter” originates from the role of materials that were used to remove the final remnants of oxygen from vacuum tubes.)

In the All-Vacancy type ReRAM, the two or more conductance states of the ReRAM are accounted for by the movement of oxygen vacancies to different parts of the device structure.

Here, the interface film, an Al2O3:Ti film, is once again represented in yellow, and the part that stores vacancies is described as a vacancy reservoir, and is designated by a yellow-orange colour.  The metal or combination of materials chosen are those that will ensure a high density of vacancies.

One important differentiating point between the HRS and LRS of these two cell types is that for the All-Vacancy ReRAM oxygen plays no role, other than its removal from normal device operation.  This means that the key distinguishing feature between the Oxygen-Ion Drift and All-Vacancy types of ReRAM, for similar sets of materials and structures, must rest with Forming and, in one case, removing the oxygen without damage to the electrode.  Examples of this damage can be found in a 2016 paper in the journal of Advanced Materials by Adnan Mehonic et al.

What are the driving forces that allow for the movement of either ions or vacancies during normal operation?  The list of possibilities is extensive, with ionic conduction and thermal diffusion high on the list, and thermophoresis, electrochemical potentials, metal valency changes, redox processes and electromigration as possible contributing factors.


A lot of brilliant minds and effort are currently being applied to understanding ReRAM operation and fabrication, and with good reason.

Outside of that scientific effort, those who hope to claim the high ground for Artificial Intelligence (AI) and neuromorphic computation, on behalf of their countries or organizations, would do well to remember that the characteristics provided by ReRAMs give us some of the keys to future, and for its realization they should provide the required support.

The general impression is this: Today’s understanding of the operation of oxide-based ReRAMs now provides a better foundation for companies like Intrinsic (UK), Weebit-Nano (Israel), and TetraMem Inc. (USA), who are working to commercialise ReRAM devices than was available in times-past for those pursuing PCMs.

It helps their cause that first-generation ReRAM arrays have been available from Fujitsu for a couple of years, with their latest product a 12Mbit memory with a Serial Peripheral Interface (SPI).  These chips provide some safe ground on which to build the ReRAM future.

One thing is very clear: those who understand, during device fabrication, how to control the defect and vacancy density in oxides, the complexities associated with the properties of oxides, and oxide-metal electrode interfaces, are likely to have more chances of success than those who do not.

This is especially important, as it appears that the control of vacancies will be important to third-generation ReRAMs, which will most likely be based on vacancy modulation of Schottky contacts.


I would like to acknowledge and thank Prof Kenyon (UCL) and Prof J. Joshua Yang, Univ. of Southern California, for some helpful correspondence and assistance.