New Embedded Memories Ahead

The embedded memory market is beginning to heat up, fueled by a new wave of microcontrollers (MCUs) and related chips that will likely require new and more capable nonvolatile memory types.

The industry is moving on several different fronts in the embedded memory landscape. On one front, traditional solutions are advancing. On another front, several vendors are positioning the next-generation memory types, such as FRAM, MRAM, ReRAM and even carbon nanotube RAMs, for the embedded market.

The new memory types have been in the works for years, and some are even shipping. Chip customers are now taking a harder look at these technologies, as the current solutions may not address all of the problems.

In one simple example, today’s MCUs integrate several components on the same chip, such as a CPU, SRAM, embedded memory and peripherals. The CPU executes the instructions. SRAM is integrated on the chip to store data or frequently used instructions. Then, embedded memory, such as EEPROM and NOR flash, are used for code storage and other functions.

Fig. 1: Standard embedded NOR flash technology. Source: Silicon Storage Technology.

SRAM and embedded memory work, but these technologies have some limitations. SRAM is fast, but it takes up too much space on the chip. Embedded memory is reliable but slow.

Chip customers will continue to use the traditional technologies for many applications, but future systems may require something different, namely the new memory types or the existing ones at more advanced nodes.

“What we are seeing in embedded memory is a movement toward valuing speed and efficiency,” said Dave Eggleston, vice president of embedded memory at GlobalFoundries. “The need is also there for versatility. By versatility, I mean beyond the usual embedded nonvolatile memory job of storing code. We also want to start to store data and possibly even state in a nonvolatile way, which helps the entire system.”

But should chip customers stay with the traditional solutions or move to the next-generation memory types for embedded apps? There is no simple answer. The traditional solutions are known quantities based on CMOS processes. But while the new memories boast some impressive specs and can even replace traditional embedded memory and/or SRAM, they also have some issues. These technologies rely on relatively exotic materials and switching mechanisms.

“That’s been a leading cause of difficulties for many of these new technologies to get off the ground,” said Jim Handy, an analyst with Objective Analysis. “The new materials add cost to the wafer, and they will often impact yield.”

To be sure, it’s a confusing landscape. To help the industry get a handle on the market, Semiconductor Engineering has taken a look at the status of the new embedded memory types and the challenges ahead.

What is embedded memory?
Next-generation memory types are targeted for various markets, such as embedded and high-end systems. For high-end systems, there is an explosion of data, causing a growing bottleneck or latency gap in systems.

To solve the problem, the industry has been searching for a new memory type that fits between DRAM and NAND in the traditional memory/storage hierarchy.

For this, the candidates include MRAM, phase-change and ReRAM. Some new memory types are beginning to ship in various parts of memory/storage hierarchy, while others are still on the runway. A few technologies will never make it—they are too expensive and difficult to manufacture.

So, it’s still too early to say where these technologies will eventually land in systems. “You are seeing many, many types of memories. You see at least 10 varieties, maybe more. People are still thinking about ways to work on certain things to solve the problems,” said Raman Achutharaman, corporate vice president and general manager of the Etch Business Unit at Applied Materials.

It’s a similar scenario in the embedded market, as the industry is still sorting out these new memory technologies. Basically, the embedded chip market involves ASICs, MCUs and system-on-a-chip (SoC) designs. These devices are embedded in a multitude of applications, such as automotive, consumer, industrial, IoT, medical and others.

Generally, MCUs incorporate SRAM-based cache memory. “Moving data on and off a chip is very slow,” Objective Analysis’ Handy said. “SRAM is used for data that changes all the time. SRAM is there just to save the amount of time it would take for the processor to communicate with DRAM.”

In addition, MCUs incorporate EEPROM or NOR flash for embedded memory. “With EEPROM, each bit is two transistors. And each byte can be erased or re-programmed,” Handy said. “On each block (with NOR flash), we have one huge transistor that does the erase for all of the bits on the block. A huge transistor still saves a lot of chip space, compared to two transistors per bit.”

Another flash memory type, NAND, is not used for embedded memory in MCUs. It’s used for storage in flash drives, SSDs and other products.

Generally, NOR dominates the embedded memory landscape. Today’s embedded NOR has 40nm feature sizes with 28nm in the works. In the lab, Renesas has fabricated embedded NOR at 16nm/14nm, but this won’t appear for several years.

“We are expecting some of the high-end IoT devices that need on-chip RF integration to take advantage of embedded flash availability at 28nm,” said Vipin Tiwari, director of worldwide marketing and business development at Silicon Storage Technology, the embedded memory subsidiary of Microchip. “We believe that 28nm will be a long-lasting node for automotive and high-end IoT applications.”

Basically, there are several types of NOR architectures, including floating gate and charge trap. In floating gate, the charge is stored in the gate. “A floating gate is made out of conducting polysilicon,” Objective Analysis’ Handy said. “With charge trap, the floating gate is an insulating layer. Hence, that makes some of the processing a little bit easier. It also allows you to make the floating gate smaller. It holds more electrons.”

Both floating gate and charge trap are widely used in the market. “It is still going to be used for a long time,” GlobalFoundries’ Eggleston said. “It does that one job really well. It stores data in a harsh environment.”

NOR has some limitations, however. It requires a byte- or sector-erase operation before the write operations. “Write speeds are very slow compared to other memory technologies. Also, the endurance is somewhat limited,” Eggleston said.

Embedded flash also may require up to 12 extra mask steps, thereby impacting cost and complexity. In addition, NOR is running into problems as it scales to 40nm and beyond. “Starting at 40nm, the characteristics and performances of NOR flash are degrading,” said Sylvain Dubois, vice president of marketing and business development at Crossbar, a ReRAM chip startup.

Going forward, customers could stay with NOR. They could also move NOR outside the MCU. But this solution would take up too much real estate, and security is an issue.

Another idea is to move to a new memory type for use in the automotive, industrial and IoT areas. “Customers have an interest in some of the other and more unique solutions,” said Walter Ng, vice president of business management at UMC. “The new memory types are making incremental progress every year. Some of these solutions are even working in niche areas.”

So what’s the best solution? “It hasn’t been resolved yet,” Ng said. It’s still too early to say which technology will prevail over the long haul. They need more time to develop, he said.

Over time, though, the industry will likely use several next-generation memory types. “I would expect these advanced memories to first find homes in applications that recognize or leverage one of their unique advantages,” said David Fried, chief technology officer at Coventor.

FRAM emerges…again
To be sure, though, the re-emergence of ferroelectric RAM (FRAM) is a big surprise. For years, FRAMs have been shipping in several embedded applications, although the technology has taken a backseat to MRAM, phase-change and ReRAM. Today, Cypress, Fujitsu and TI are shipping FRAM-based chips.

Using a ferroelectric capacitor to store data, FRAM is a nonvolatile memory with unlimited endurance. FRAM is faster than EEPROM and flash. FRAM performs an over-write function in the memory cell without an erase operation.

Basically, an FRAM cell consists of a crystal structure, based on lead zirconate titanate (PZT). In simple terms, the outer portion of the crystal is lead. By applying an electric field, zirconate/titanate ions move inside the crystal. Then, the capacitor plots the polarization of the ions. And finally, data is stored in the form of “1” or “0”.

FRAM has some pluses and minuses. “Ferroelectric could be written with very low power, like three orders of magnitude of less energy than is required by NOR flash or EEPROM for a write cycle,” Objective Analysis’ Handy said. “But lead ions are very mobile. They move around within the silicon. That contaminates your process.”

FRAMs are also limited in terms of scaling. To solve these and other problems, NaMLab and a spin-off company, Ferroelectric Memory (FMC), are pioneering the development of a next-generation FRAM, dubbed a ferroelectric FET (FeFET).

A FeFET makes use of ferroelectric properties in thin doped hafnium oxide layers. “In general, you replace the conventional logic gate dielectric with a ferroelectric material, a dielectric that remembers the electric field to which it had been exposed,” said Stefan Müller, chief executive of FMC. “In FeFETs, a permanent dipole is formed within the gate dielectric itself, splitting the threshold voltage of the ferroelectric transistor into two stable states. Accordingly, binary states can be stored in the FeFETs similar to how it is done in a flash memory cell.”

FMC, NaMLab and GlobalFoundries are working on FeFETs for the embedded memory market. In the lab, GlobalFoundries devised a one-transistor FeFET technology into its 28nm CMOS logic process. FeFET is also ideal for an embedded memory solution within GlobalFoundries’ 22nm FD-SOI platform, according to Martin Trentzsch, senior section manager for technology and integration at GlobalFoundries.

Fig. 2: Transmission electron microscopy image in dark filed mode of a ferroelectric HfO2-based capacitor structure. Single HfO2 grains with similar crystallographic orientations are visible. Source: NaMLab.

Using only two extra mask steps, GlobalFoundries devised a 64-kbit FeFET test chip. “It’s a low-power, cost-effective solution for code storage,” Trentzsch said. “You can theoretically put these transistors and embed them in regular logic with standard logic design rules.”

MRAM and ReRAM
Besides FeFET, GlobalFoundries is also developing another embedded memory technology type–MRAM. Intel, Samsung, TSMC and many others are also developing MRAM.

Offered as a standalone device or as an embedded memory solution, MRAM delivers the speed of SRAM and the non-volatility of flash with unlimited endurance. The industry is currently developing a second-generation MRAM technology called spin-transfer torque MRAM (STT-MRAM or ST-MRAM).

In STT-MRAM, the storage element consists of a one transistor, one magnetic tunnel junction (MTJ) memory cell. “When a bias is applied to the MTJ, electrons that are spin polarized by the magnetic layers traverse the dielectric barrier through a process known as tunneling,” according to Everspin, an MRAM supplier.

Everspin is shipping a standalone ST-MRAM using a perpendicular magnetic tunnel junction (pMTJ) technology. This is 256-Mbit DDR3 product. A 1-gigabit ST-MRAM is in the works, which represents a breakthrough. “With perpendicular MTJ, (ST-MRAM) can be scaled,” said Er-Xuan Ping, managing director of memory and materials within the Silicon Systems Group at Applied Materials, in a recent interview.

Everspin is also working on an embedded version of the technology. Designed to replace EEPROM, flash and SRAM, Everspin’s embedded MRAM technology is integrated in the last two metal layers of a standard CMOS logic processes using just a few mask steps.

Generally, STT-MRAMs are challenging to make, however. “There are still numerous obstacles for developing MRAM products, such as MTJ short fail, relatively narrow sensing margin and MTJ patterning difficulties,” said Y. J. Song, a principal engineer with Samsung.

Basically, an MTJ consists of a thin barrier layer of magnesium oxide (MgO), which is sandwiched by two ferromagnetic layers based on a cobalt-iron-boron (CoFeB) compound. The barrier layer is only ≤1nm.

“This structure involves new materials that are nonvolatile,” said Yang Pan, chief technology officer for the Global Products Group at Lam Research. “For MRAM, the challenge is how to define this structure.”

In the fab flow, the CoFeB/MgO/CoFeB stack is formed using deposition. Then, the stack is etched, forming the MTJ cell. Problems could surface if these steps aren’t precise. Shorts can occur, thereby impacting yield. “If you use conventional etch technology, these materials reactively sputter and then re-deposit on the sidewall of the structure. That re-deposition shorts out the device. It’s a killer,” added Rick Gottscho, executive vice president of global products at Lam.

To solve the problem, Samsung has devised several new manufacturing techniques. For example, instead of a traditional etch process, Samsung has devised an ion beam etching (IBE) technology, enabling it to reduce the short fails below 1 ppm.

Using IBE and other techniques, Samsung has fabricated an embedded 8-Mbit STT-MRAM within its 28nm process. “(The device proves the) feasibility of eMRAM commercialization for IoT applications,” Samsung’s Song said.

For its part, GlobalFoundries plans to offer embedded MRAM within its 22nm FD-SOI platform. In one example of an application, an MCU could incorporate embedded MRAM and SRAM. MRAM would replace embedded flash for code storage.

Embedded MRAM could also assume some of the SRAM-based cache functions, thereby saving space and cost. “You don’t get rid of SRAM, but you might reduce the amount of SRAM you have on board,” GlobalFoundries’ Eggleston said. “The SRAM and eMRAM work together.”

Another technology, ReRAM, is nonvolatile and based on the electronic switching of a resistor element material between two stable resistive states. ReRAM delivers fast write times with more endurance than today’s flash.

Many are targeting ReRAMs as a replacement for NAND, although ReRAM is also geared for embedded apps. For example, China’s Semiconductor Manufacturing International Corp. (SMIC) is developing 40nm embedded ReRAM based on the technology from Crossbar. This will enable embedded ReRAM for MCUs and SoCs.

ReRAM is ideal for applications requiring low-power code execution and data logging from sensors. “Traditional flash technology requires high voltage programming and slow block erase commands before programming the memory again,” Crossbar’s Dubois said. “ReRAM does not require high voltage programming and can be written over without any erase commands. The tasks are performed faster at lower power.”

Developing ReRAMs with good cycling and stable retention at high temperatures is challenging, however. Cost, of course, remains an issue for both ReRAM and MRAM.

Nanotube RAMs
For years, the industry has been talking about carbon nanotube FETs for logic. Carbon nanotubes are cylindrical structures, which are strong and conductive.

One company, Nantero, is developing carbon nanotube RAMs, dubbed NRAMs. “It’s a different application,” said Greg Schmergel, chairman and chief executive of Nantero. “It’s as fast as DRAM, while being nonvolatile. We do have unlimited endurance as well.”

NRAMs are targeted for standalone and embedded. “It’s based on carbon nanotubes, which are either in contact with each other or not in contact with each other to form high resistive and low resistive states. So you have very distinct ‘0s’ and ‘1s’. This is because there is a large difference between the ‘on’ and ‘off’ state,” Schmergel said.

Fujitsu is making NRAMs on a foundry basis for Nantero. In 2018, Fujitsu is also expected to offer the first NRAM products based on 55nm technology with 40nm in the works.

While suppliers of the new memory types are making progress, they still face an uphill battle. Cost, integration and manufacturability are just some of the issues.

And over the years, suppliers have made bold claims, but many have failed to deliver. “There has always been skepticism,” Schmergel said. “But the openness for new options has gone up dramatically. The industry has really started to think about what (these new technologies) could do for their products.”

Related Stories
Sorting Out Next-Gen Memory
A long list of new memory types is hitting the market, but which ones will be successful isn’t clear yet.
Tech Talk: Embedded Memories
Which type of memory is better and why.
The Future Of Memory (Part 1)
Experts at the table, part 1: DDR5 spec being defined; new SRAM under development.