What Ever Happened To Next-Gen Ferroelectric Memories?

FeFETs, FeRAM and other next-generation ferroelectric memories are in the works, but it’s unclear when they will appear.

By Mark LaPedus

For years, several semiconductor companies, R&D organizations and universities have been working on next-generation ferroelectric memory devices, a technology that was once supposed to shake up the market.     

So far, though, next-generation ferroelectric memories, such as the long-awaited FeFET, have not been commercialized or put into production amid ongoing technical and economic challenges with the technology. Originally, the initial devices were expected to reach the qualification stage around 2019, but the technology is still a work in progress.  

Still in R&D, the new ferroelectric memories include a family of memory types, such the FeFET, FeRAM, FeCAP, and embedded ferroelectrics, according to the Georgia Institute of Technology. All of these technologies leverage ferroelectric properties to store data in devices. But each memory type is different and in various stages of development. (See chart below)    

Behind the scenes, leading memory companies and foundry vendors continue to pursue or explore one or more of these newfangled memory technologies, and for good reason. These devices are supposedly fast and non-volatile with low power and high endurance. They are also compatible with today’s semiconductor processes. A non-volatile memory retains stored data even when a system is shut off.

Not long ago, some believed that next-generation ferroelectric memories could potentially displace the existing memory types in the market, namely DRAM and NAND flash, which have some limitations. But as it stands today, the new ferroelectric memories won’t replace existing memory for the foreseeable future, if ever. It’s difficult to dislodge existing memory. Still, if they ever move into production, the newfangled ferroelectric memories could carve out a new market, driving a wave of applications.   

Next-generation ferroelectric technology emerged over a decade ago, when researchers stumbled upon the effects of ferroelectricity in hafnium oxide (HfO2), a common material used in chips. Later, researchers developed a new technology called a ferroelectric field-effect transistor (FeFET).

FeFETs and related technologies have generated interest, and confusion, in the market. They are sometimes confused with ferroelectric RAMs (FRAMs). Commercialized in the 1990s, FRAMs are an older ferroelectric memory type used for niche applications. Nonetheless, FeFETs and related technologies are different than FRAMs.

In one configuration, a FeFET resembles a traditional metal-oxide-semiconductor field-effect transistor (MOSFET). A transistor is a tiny structure that amplifies or switches electrical signals inside a chip. Today’s advanced chips incorporate billions of tiny transistors in the same device.      

A logic-based MOSFET is a planar device with a source and a drain structure on each end of a substrate. A gate structure resides in the middle. In operation, a gate voltage establishes a field. This in turn allows or blocks a current flow between the source and drain.       

To make a FeFET, HfO2 ferroelectric materials are integrated in the gate stack of a logic transistor. In doing so, the logic transistor is transformed into non-volatile FeFET memory transistor. Compared to existing memory, FeFETs utilize a different mechanism to store data. “The oxygen atoms of HfO2 can reside in two stable positions, shifting either up or down according to the polarity of an externally applied electric field,” according to Ferroelectric Memory Company, a developer of ferroelectric memories. “Therefore and depending on the position of the oxygen atoms, a permanent electric dipole is created that can either point upwards or downwards in this way, enabling the storage of two binary states.”

(Structure of a 1-transistor FeRAM cell. FeRAMs and FeFETs are among the next-generation ferroelectric memory types. Source: Cyferz at English Wikipedia)

Clearly, the new ferroelectric memories are complex with several technical challenges. The endurance is still a question. Putting them in production is another issue. “The materials are compatible with silicon manufacturing. There is no problem with that. But mass production requires variability control. Of course, the application/market needs to be further reviewed,” said Shimeng Yu, Dean's Professor of the School of Electrical and Computer Engineering at the Georgia Institute of Technology. 

Indeed, there are some economic issues. A company could spend millions of dollars to develop a new memory, but there’s no guarantee that it will succeed. There are numerous examples where companies have tried but failed to commercialize a new memory type in the risk-adverse semiconductor industry.       

Still, many entities are pursuing new ferroelectric memories. Others have put it on the backburner. Here’s just some of the major events in the arena:  

*Micron recently presented a paper on an NVDRAM, a 3D ferroelectric memory.

*Kioxia is working on what appears to be a ferroelectric-based 3D NAND-like technology.

*Imec and Georgia Tech are working on FeCAPs, a non-volatile ferroelectric capacitor technology.

*GlobalFoundries and Intel are working on an embedded memory technology based on ferroelectrics. To date, GlobalFoundries is the only company that has shipped next-generation ferroelectric memory, mostly for R&D.        

*Samsung, TSMC and others are working on it.

(Next-gen ferroelectric memory types. Source: Georgia Institute of Technology)

DRAM, NAND challenges

Today’s computers, smartphones and other systems have the same basic architecture. They incorporate a circuit board. A central processing unit (CPU), memory, storage and other components reside on the board. 

The CPU handles the data processing functions. A system also requires different types of memory chips. The first type, called SRAM, is integrated with the CPU. SRAM, known as cache memory, stores frequently accessed data.      

DRAM, another memory chip type, handles the main memory functions in systems. These chips enable the CPU to access short-term data at high speeds. In PCs, individual DRAMs are bundled in a module.

To store the data, a system also incorporates either a solid-state drive (SSD) or a hard-disk drive. SSDs incorporate NAND flash memory chips. In operation, data is written, read, and erased on a NAND chip. “Each chip contains one or more dies, and each die contains one or more planes. A plane is divided into blocks, and a block is divided into pages,” according to a blog from Redgate Software, a data management software supplier.

Data is read and written at the page level, but erased at the block level. “A page is made up of multiple cells that each hold one or more data bits. Each bit is registered as either charged (0) or not charged (1), providing the binary formula needed to represent the data,” according to the blog.

For years, DRAMs and NAND flash have been the mainstream memory devices in systems. They are inexpensive and reliable. However, DRAM and NAND flash memory also have some limitations. For example, NAND flash is non-volatile. But NAND is slow and the endurance is sometimes an issue.

A DRAM is fast, but it’s volatile, meaning the data is lost in the memory device when the system is shut down. And in operation, DRAMs must be refreshed every few milliseconds to retain data, which elevates the power consumption in systems. This refresh process can also cause latency in systems.  

That’s not the only problem. Back in the early 2010s, NAND flash memory was running out of steam. Scaling the DRAM was also challenging.

Even before then, many saw the inherent problems with DRAM and NAND. That began a frantic search to develop a new memory type. The goal was to develop a new single memory that combined the non-volatility of flash with the speed of DRAM. It had to be inexpensive with low power and high endurance.   

New memory race

Over time, semiconductor companies and R&D organizations introduced so-called emerging or next-generation memories. These new memory types included carbon nanotube RAMs, FRAMs, MRAMs, phase-change memory (PCM) and ReRAMs. Next-generation ferroelectric memory also emerged.    

A decade ago, many of these new memory types were touted as replacements for DRAM or NAND—or both. Some used the term “universal memory,” which described a single memory that could do everything. 

But the death of DRAM and NAND was greatly exaggerated. Over the years, suppliers of DRAM and NAND found new and innovative ways to extend these technologies. Both remain the mainstream memory technologies for today’s systems.

“You really only need two types of memory at this point, not including cache/SRAM. For 90% of the work, you need fast DRAM for random memory and NAND for storage. Hard drive is the other part of storage. That’s really all you need,” said Mark Webb, a principal/consultant at MKW Ventures Consulting. “We already have DRAM, which is fast and you can cycle it forever. Then, we have NAND, which is inexpensive. Few care about the fact that NAND isn't fast and DRAM is volatile.”

Meanwhile, over the years, the new memory types were unable to keep up with the advancements and price points with both DRAM and NAND. All told, the emerging memories never live up to their past promises and didn’t displace DRAM and NAND. They weren’t a total bust, however. A few new memory types have experienced some success. Some are just getting off the ground, while others are still in R&D. A few simply failed.

Some new memory types carved out a niche that falls somewhere between DRAM and NAND. Some call this segment storage-class memory. “The people that do want something in between DRAM and NAND, it’s usually a low-volume product. But they will use it, because it may fit an application. That’s the way design works,” Webb said.

Nonetheless, each new memory type is different. PCM, one emerging memory type, is viewed as both a limited success as well as a spectacular failure. Conceived back in the 1960s, PCM uses different material sets and mechanisms to store data. In 2015, Intel introduced Optane, a next-generation non-volatile memory chip line based on PCM technology. Intel shipped Optane devices in the market, but it failed to gain widespread adoption. Intel lost millions of dollars with Optane, causing it to drop the product in 2022.

Micron sold a similar device, but the company also exited the PCM market. PCM is still around, however. STMicroelectronics is pushing PCM for embedded memory applications. (This article will address embedded memory later.) 

Other new memory types also flopped. For years, Nantero was developing carbon nanotube memories, a non-volatile technology that claimed to be faster and denser than DRAM. Nanotero struggled to put these memories into production. Unable to obtain funding last year, the company recently went under and no longer exists.    

ReRAM, another emerging memory type, has been in the works for years. ReRAM is a non-volatile memory “that works by changing the resistance across a dielectric solid-state material rather than directly storing a charge,” according to Weebit Nano, a developer of ReRAM. “ReRAM typically has 10x-100x better endurance than flash, handling between 100,000 and a million write cycles versus the typical 10,000 program/erase cycles that flash can manage.”

To date, though, ReRAM has experienced limited success. A few companies have shipped or are sampling ReRAM devices, while others are in R&D. TSMC offers a ReRAM process for embedded memory applications.

Of all the next-generation memory types, MRAM has seen the most success. For years, Everspin has been shipping MRAM devices for various applications. Everspin’s most advanced part is a 1Gb device, with a 2Gb technology in R&D, according to the company’s roadmap. 

The most advanced MRAM technology is called spin-transfer torque (STT) MRAM. STT MRAM is non-volatile and works by using “the spin-transfer torque property, which is the manipulation of the spin of electrons with a polarizing current, to establish the desired magnetic state of a magnetic tunnel junction (MTJ),” according to Everspin.

Embedded memory options

Everspin sells standalone MRAM chips. Others sell standalone ReRAM and FRAM devices. Perhaps the biggest opportunity for the next-generation memory types is in the embedded memory space. STT MRAM is making inroads here, and so are FRAM, PCM and ReRAM.

Embedded memory is a simple concept. The idea is that you embed a tiny memory in a chip. The microcontroller (MCU) is one example here. An MCU is a chip that is used to power and control a system. MCUs are used in appliances, cars, industrial equipment, medical products and other systems.

Typically, an MCU integrates a CPU, memory and other chips on the same device. In an MCU, the memory is called embedded memory. And in many cases, NOR flash memory is the technology of choice for embedded memory. NOR flash memory, a non-volatile technology, is related to NAND flash memory. But in many ways, NOR and NAND flash memory are different. NOR is fast and is typically used for code storage in systems.

MCUs, along with the embedded memory, come in different configurations. They are manufactured by semiconductor companies in large facilities called fabs. In a fab, MCUs (and embedded memory) are manufactured using various process technologies. A process technology, sometimes called a process node, “refers to a specific semiconductor manufacturing process and its design rules,” according to WikiChip, a technology site.

In the semiconductor world, there are different process nodes and each one is represented by a number. A lower node number means that the process, and the chip product, are more advanced. 

Today, the more advanced MCUs are manufactured using a so-called 28nm process node. These MCUs incorporate embedded NOR flash memory, also based on a 28nm process. Last year, Renesas introduced the first MCUs (with embedded memory) based on a more advanced 22nm process.  

At the 28nm and 22nm process nodes, NOR flash memory becomes more difficult and expensive to manufacture in a fab. Simply put, NOR flash is nearing its physical limits. “All flash, NAND and NOR, runs out of steam at ~15nm,” said Jim Handy, an analyst from Objective Analysis.

In the past, a NAND flash memory chip was a lateral or 2D structure, where the memory cells were placed side by side. A decade ago, traditional 2D NAND did indeed run out of steam around the 15nm node.

In response, NAND flash memory suppliers found a new solution. They moved from a 2D to a 3D structure. Starting in 2013, suppliers introduced 3D NAND. In 3D NAND, the flash memory cells are stacked vertically, which increases the storage density.

NOR isn’t so lucky. “When CMOS logic jumped from 28nm to 14nm finFETs, they crossed the 15nm threshold, so there’s no NOR at and beyond 14nm,” Handy said.   

That’s where emerging memories fit in. They are targeted to replace NOR flash for embedded memory applications at the 28nm/22nm nodes and below. “This is the area that MRAM, ReRAM, FRAM, and even PCM hope to take over,” Handy said. (Handy has co-authored a new report on emerging memories.)

Several foundry vendors, such as GlobalFoundries, Samsung, TSMC and others, offer STT MRAM processes for embedded memory applications. TSMC offers an embedded ReRAM process. STMicroelectronics uses embedded PCM for some of its MCU lines.

Basically, foundry vendors manufacture chips for other companies in large fabs. In one example, an MCU supplier would select a foundry vendor to manufacture its chip lines. In turn, the foundry vendor would manufacture the MCU, along with the embedded memory, in the same process. This is a simple explanation for a complex process flow.      

What happened to ferroelectrics?

Years ago, ferroelectrics were supposed to be the next big thing. In the 1990s, several companies introduced the original ferroelectric RAMs, or FRAMs. Infineon and Fujitsu sell standalone FRAM chips. FRAM can be used as an embedded memory. As stated, FRAMs are different than FeFETs.

“FRAMs are still shipping, and they are still very popular because of their fast write, their extremely low write energy, and their immunity to radiation for deep space applications,” Objective Analysis’ Handy said.

FRAMs have some limitations. These devices can’t scale beyond a certain point. For that reason and others, FRAMs became a niche technology and fell off the radar screen.

Then, FeFETs, which are also ferroelectric, emerged and created a buzz. There are some major differences between FRAMs and next-generation ferroelectric technologies like FeFETs. “Most of the older FRAMs use 2T2C cells (two transistors and two capacitors), which are the ferroelectric part. That’s a pretty big cell,” Handy said.

“FeFETs got rid of one transistor and both capacitors. That’s an impressive difference. In essence, a FeFET has the potential to be as dense as flash, either NOR or NAND, implying that it can be as inexpensive as NOR or NAND. Ferroelectric memories are far faster to program, and they don’t require erase-before-write. They also need far less power to write, and they don’t require the energy-squandering refresh cycles that DRAM does,” he said. “So, you come away with the possibility that they could become as cheap as flash, while slashing power consumption in servers and battery usage in portable devices.”

Potentially, ferroelectrics could also enable new devices with DRAM-like performance. This is accomplished by deploying ferroelectric materials as the dielectric in a DRAM capacitor. “You could actually put it in a DRAM process with less than 10% of the process steps being changed, and you could ramp it into high-volume in an existing fab. Now that's just the theory,” said MKW’s Webb. “If it works out that you could just implement that on a DRAM process, that would be a game changer.”

Meanwhile, FeFETs are just one of many types of next-generation ferroelectric memories. The others include FeRAM, FeCAP or nvCap, BEOL ferro, and FEOL ferro, according to Georgia Tech.

There are two flavors of FeFETs. The first is a logic-compatible FeFET. The other is ferroelectric-based vertical device for storage.

At the recent VLSI Symposium, Kioxia presented a paper on a dual-layer 4F2 vertical device. The device is a 30nm-diameter vertical channel-all-around FeFET for high-speed memory applications. The key component is a titanium dioxide (TiO2) channel structure, which has better endurance than silicon-channel FeFETs.

Also at VLSI, Samsung demonstrated a novel FeFET with an IGZO (indium-gallium-zinc-oxide) channel, featuring a record-high memory window of up to 17.8V. This is targeted for non-volatile memory applications.

Meanwhile, at last year’s IEDM conference, Micron presented a paper on a 3D FeRAM device called an NVDRAM. The device resembles a DRAM with non-volatility. It’s a stacked 32Gb device with near-DRAM performance.

However, Micron, according to multiple sources, has put the NVDRAM on the backburner or elected to not pursue it. After its ill-fated experience with PCM, Micron may have felt that it’s too risky to bring up another new memory type. Plus, the NVDRAM may cannibalized its profitable DRAM business.  

Georgia Tech, meanwhile, has been working on a different ferroelectric technology called an nvCaP. Some call it a FeCAP or ferroelectric capacitor. “A FeCAP resembles a traditional capacitor but has a ferroelectric material in between two metal layers (the electrodes) instead of a conventional dielectric material,” said Jan Van Houdt, a fellow at Imec and program director of ferroelectrics at the R&D organization, in a recent blog.

Georgia Tech is exploring the idea of using FeCAPs in compute-in-memory (CiM) applications for deep learning, which is a form of AI. The FeCAPs would be used to store the weights in analog CiM applications.

There’s just one problem. Hafnium-zirconium oxide (HZO), a ferroelectric material, is used for the FeCAP structure. But the scheme to read the polarization state of HZO-based FeCAPs is destructive, which presents a stumbling block for CiM applications, according to Imec.

Imec and Georgia Tech may have found a solution. They recently presented a paper on a mechanism that enables a non-destructive read operation. “This is a new mechanism to operate ferro using small-signal C-V asymmetry (sense high and low capacitance) for a non-destructive read,” Georgia Tech’s Yu said.

Meanwhile, at the VLSI Symposium, Intel presented a paper on low-voltage FeRAM capacitors for embedded memory applications. The FeRAM capacitors have been demonstrated with write PL/BL voltage scaled down to 1.3/1V, and a read voltage scaled down to 0.6V with an extrapolated 10-year fatigue/breakdown reliability at 85 degrees C.

Another company, GlobalFoundries, is working on a 28nm/22nm logic-compatible FeFET process for embedded memory applications. The process utilizes fully-depleted silicon-on-insulator (FDSOI) technology, which is targeted to reduce the power. GlobalFoundries, according to sources, has sold some FeFET wafers to universities. It’s still in R&D.

Conclusion

Over the years, a number of entities have presented a plethora of papers on ferroelectrics. It’s impressive work.  

Still, the big question is clear: Will FeFETs and the next-generation ferroelectric memories ever become commercialized and put into production? That’s still unclear.

“If you want new memory, there are several of those in the market (MRAM, PCM, ReRAM). So do we need yet another new memory? If ferroelectric wants to take off, it has to be better than other new memories,” said MKW’s Webb.