What is the future of memory? MRAM is most promising

What is the future of memory? Intel is undoubtedly one of the manufacturers that has bet the most on future technologies. From Rambus DRAM to later Optane 3DXpoint memory, it has shown its exploration and desire for more advanced technologies. Unfortunately, these two memory technologies that Intel had high hopes for were ultimately sentenced to death by their own hands, which makes people sigh.

And after the curtain call of Optane, what technologies are worthy of everyone's attention?

Tom Coughlin from Coughlin Associates and Jim Handy from Objective Analysis recently released a report. Two semiconductor analysts conducted a detailed analysis of the prospects of five emerging storage technologies, from which we may be able to get a glimpse of technology development.

Analysts first summarized the lessons learned from Optane's failure. The essence of semiconductor manufacturing is that the higher the output, the lower the cost. With Optane, Intel could have increased production capacity to reduce prices and drive chip sales. However, Optane's initial production capacity was not sufficient, which meant that the cost of chips was higher, and it had to bear this part of the loss itself. Sales must continue to increase until the increase in production capacity is justified, and ultimately the cost of each chip is reduced, thereby making a substantial profit.

This also shows that economies of scale may play a greater role in the emerging memory market than we think, and the report provides a conclusion that the wafer volume must be close to 10% of the volume of competing technologies to achieve cost parity.

In the process of Optane's gradual failure, five emerging storage technologies have begun to appear on the stage, including MRAM, phase change memory (PCM), ferroelectric RAM (FERAM), resistive memory RAM (ReRAM) and NRAM/UltraRAM. They are expected to surpass the expansion limitations of NAND and NOR and consume less power than DRAM and SRAM.

FRAM/FeRAM

FRAM, invented in 1952, is the oldest emerging memory. Today, more than 4 billion FRAM chips are installed in various devices. Although it has iron in its name, FRAM does not use any iron. It just has a hysteresis loop similar to ferromagnetism, and this hysteresis loop allows it to store data.

The principle of FRAM is to exploit the unique physical properties of certain crystal lattices. In ferroelectric materials, atoms can occupy one of two stable positions within the lattice. The electric field moves the mobile atoms within the lattice to one of the two stable positions, depending on the polarity of the electric field and some physical property (perhaps capacitance or resistance), depending on the position of the trapped atom.

There are currently many manufacturers still producing FRAM. For example, Infineon mainly produces discrete FRAM chips, while Texas Instruments and Fujitsu embed the technology into MCUs. Fujitsu also embeds FRAM in subway tickets. The main reason for this use is that the writing energy consumption of FRAM is relatively the lowest among storage technologies.

Why did it take so long for FRAM to be invented, why is it still unknown after billions of chips have been produced, and is it still listed as an emerging storage technology?

The reason is that FRAM was previously mainly based on lead zirconate titanate (PZT) and strontium bismuth tantalate (SBT), but both materials contain lead or bismuth, which will cause pollution to the fab, thus limiting its production capacity. Fortunately, in 2011, it was discovered that hafnium oxide (HfO) has ferroelectric properties under certain conditions. HfO is the basis of the high-K gate dielectric used in FinFET. It not only solves the production capacity problem, but also does not cause pollution. Therefore, although HfO has not been officially used for production, the future prospects are very bright.

Compared with flash memory, the advantages of FRAM include lower power consumption, faster write speeds, and higher maximum read/write endurance. FRAM has a data retention time of more than 10 years at +85°C (up to decades at lower temperatures), but it also has its own disadvantages, that is, the storage density is much lower than flash memory devices, the storage capacity is limited, and the cost is higher. As of 2021, the storage size (density) of chips sold by different vendors does not exceed 16Mb.

Currently, FRAM is now being embedded into chips via CMOS technology, allowing MCUs to have their own FRAM memory, which is fewer stages than is required to embed flash memory in MCU chips, thus significantly reducing costs.

PCM

Due to Intel's launch of Optane memory, phase change memory (PCM or PRAM) has long been the revenue leader among emerging memory technologies. In fact, as early as 1970, Intel's Gordon Moore, Ron Neale, and D-L Nelson co-wrote an article about the 256-bit PCM prototype. Its research and development history is long and is not inferior to other storage technologies.

The origin of PCM can be traced back to 1960, when Ovshinsky established the Energy Conversion Laboratory to study amorphous materials and their phase change characteristics. The laboratory was renamed Energy Conversion Devices (ECD) in 1964, and one of Ovshinsky's many innovations was the Ovonics phase-change memory named after him. Intel eventually cooperated with ECD to obtain the intellectual property license for Ovonics phase change memory, and officially released 3DXPointPCM in 2015.

In addition to Intel, STMicroelectronics has produced microcontrollers (MCUs) with PCM program memory, and storage manufacturers such as Samsung and Micron also mass-produced PCMNOR flash memory replacement products more than ten years ago, but the existence of these products was quite short.

The basis of PCM is a glass material deposited on top of a standard CMOS logic chip. This material changes its state according to the characteristics of the glass. The glass changes from a crystalline state to an amorphous state, corresponding to a conductive or resistive state respectively. There are two ways to increase storage capacity: one is three-dimensional stacking, which is the focus of Intel and Micron, and the other is multi-value technology, in which IBM has made breakthrough progress.

Compared with flash memory, PCM has many advantages, such as strong embeddability, excellent repeatability, good stability, and compatibility with CMOS processes. In fact, so far, no clear physical limit has been found for PCM. When the thickness of the phase change material is reduced to 2nm, the device can still undergo phase change.

The biggest advantage of PCM is that it can use a cross-point configuration to store data at the intersection of two orthogonal conductive lines, which facilitates stacking, thus making the chip size and production cost lower than any mature technology except 3D NAND.

But PCM also has shortcomings that cannot be ignored. Heat is still a major problem. Although the memory is thermally stable and can handle high-temperature applications, the heat generated when programming a cell may affect its neighboring cells. Local heating can cause gaps above the battery. In addition, flash memory's ability to store and detect multiple bits per cell gives it a storage capacity advantage over PCM.

In recent years, there has been strong interest in the application of PCM in in-memory computing. The idea is to perform computing tasks, such as matrix-vector multiplication operations, in the memory array itself by leveraging PCM's analog storage capabilities and Kirchhoff's circuit laws. In 2021, IBM released a mature memory computing core based on multi-level PCM integrated in the 14nm CMOS technology node.

MRAM

Magnetic RAM (MRAM) is a technology based on the physical principles of all magnetic recording (hard disks, tapes, etc.), but its application method removes mechanical elements. As of now, Everspin, a company spawned by the research results of Motorola and Freescale, is the leader of this technology, with operating income of US$44 million in 2021.

In addition, Avalanche and Numem have recently joined the ranks of producing MRAM, and foundries such as TSMC, GlobalFoundries and Samsung have launched embedded MRAM processes. Currently, MRAM processes have begun to be used in SoCs for IoT applications and micro-power devices.

There are many types of MRAM, but their structures are very similar. They all use cobalt and magnesium layers as a combination of giant magnetoresistance (GMR) sensors and magnetic switching elements. They are also widely used in hard disk read/write heads. Their main advantage is speed. Many people have imagined that MRAM can replace high-speed SRAM in the future.

After years of research, MRAM has been divided into various types and routes: STT-MRAM effectively solves the problem of SRAM memory "leaking" energy when inactive; SOT-MRAM significantly improves the device's durability and read stability, eliminating the inherent openness in STT-MRAM devices. Off delay; VCMA-MRAM further reduces the power consumption of STT-MRAM, but the writing speed is relatively slow; VG-SOT combines the advantages of the first two, but the manufacturing process is more complex and the function needs to be verified; (VG-)SOTMRAM has greater potential in simulated memory computing...

Over the years, different types of MRAM memory devices have emerged, with trade-offs between write speed, reliability, power consumption and area consumption, with completely different applications depending on specific characteristics, such as STT-MRAM for embedded flash and last-level cache, SOT-MRAM for lower-level cache, VCMA-MRAM for ultra-low-power applications, and finally VG-MRAM. VG-SOTMRAM serves as the ultimate unified cache and also has the advantages of in-memory computing.

In MRAM, data is typically stored in a "free" layer whose magnetism can be changed and compared to a "fixed" layer set during production, and the GMR sensor is responsible for detecting the difference between the two. The biggest difference between most MRAM variants is how data is written. All MRAM uses at least one transistor per bit cell, while many MRAM uses two transistors and draws considerable current, making the technology less cost-effective to produce than other technologies.

MRAM has SRAM-compatible read/write cycles, making it particularly suitable for applications that must store and retrieve data with minimal latency. It successfully combines low latency, low power consumption, infinite persistence, scalability and non-volatility.

As a magnetic technology, MRAM is inherently radiation-resistant, making it popular for aerospace applications that are also less price-sensitive. Additionally, MRAM has found a place in enterprise storage, such as IBM's flash memory core modules, where Everspin's MRAM is used as a buffer in the event of unexpected power outages.

MRAM also has broad prospects in industrial applications. Analysts said that industrial applications require very fast writing capabilities and require non-volatile storage. However, NAND flash memory, NOR flash memory and EEPROM are all very slow to write and consume a lot of power. SRAM with additional batteries needs to replace the battery every few years. In contrast, MRAM seems to be at home in these scenarios.

The automotive industry is one of the important reasons why MRAM is so popular. Due to the increasing demand for MCUs and the rising cost of flash memory, many suppliers have begun to switch from flash memory to eMRAM. In 2022, Renesas Electronics announced the launch of STT-MRAM test chips. It said that compared with flash memory manufactured with FEOL, MRAM manufactured with BEOL has advantages in processes below 22nm because it is compatible with existing CMOS logic process technology and requires less additional mask layers.

IBM is more optimistic. "In about three years, you will be able to point to every new car on the street and say that car has eMRAM in it," said Daniel Worledge, distinguished researcher and senior manager at IBM. "There is no longer embedded flash in the advanced nodes, all the foundries have stopped developing it, and the transition period is 22nm and 28nm, depending on the foundry."

ReRAM/RRAM

In 1971, Leon Chua of the University of California, Berkeley, wrote a theoretical paper titled "Memristor—The Missing Circuit Element." The paper describes a fourth basic passive electronic device, a memristor, which can regulate the current flowing through itself based on the amount of charge previously flowing through the device. At this point, memristor is just a theory, a hypothetical device that satisfies the symmetry requirements of the equations that describe the behavior of three other basic passive electronic components: resistors, capacitors, and inductors.

Nearly forty years later, in 2008, HP Labs announced that it had successfully created a memristor using titanium dioxide. Memristor is a non-binary device that can be used to store analog or digital data. At that time, some people predicted that DRAM was about to die and memristor would be replaced in the form of resistive memory or RRAM. At that time, HP stated that it would use RRAM in its upcoming lunar computer.

But in 2015, HP reversed its decision and said it would use DRAM instead of memristors in its lunar computer. Fifteen years after HP announced the success of memristor production, the RRAM revolution has still not happened, and it doesn't seem to be happening anytime soon.

Like MRAM, there are several variations of resistive RAM (ReRAM or RRAM), which are all made by depositing special materials on top of standard CMOS logic.

The ReRAM foundry process is mainly supported by TSMC, Winbond and GlobalFoundries. Renesas (through the acquisition of Adesto), Fujitsu, Microchip and Sony produce ReRAM as independent products. Nuvoton Technology uses it in microcontrollers. Currently, many companies around the world are developing ReRAM processes.

The technical principle of ReRAM is that in a resistive RAM cell, electric current is passed through two wires to detect whether the resistance of the bit cell is high or low. Typically, the state of the cell is changed by increasing the voltage in the positive or negative direction, thereby increasing or decreasing the resistance of the cell. This is achieved by moving conductive elements such as metal ions or oxygen vacancies into the bridge, or removing these elements from the existing bridge. One might argue that most other emerging memory technologies (PCM, MRAM, and FRAM) fall into the ReRAM category because they also use variable resistors to indicate the state of memory bits.

The key feature of ReRAM is that, like PCM, it can be built into crosspoint cells for stacking, and because linear values can be stored on a single bit cell, it may also be used in neural networks in the future.

The main advantage of ReRAM is that it consumes less power and does not need to consume a lot of energy to maintain the storage state like traditional storage devices. Some RRAM materials also have multiple resistance states, making it possible to store multiple bits of data in one storage unit, thereby increasing storage density. However, it does not have advantages over other emerging technologies in terms of random read and write speed and durability.

Over the years, the number of patent applications related to ReRAM technology has been increasing. Especially after 2010, the number of patent applications has increased significantly. Samsung currently has the largest number of related patents, followed by Micron and SK Hynix. Major storage manufacturers have expressed interest in this technology.

NRAM/UltraRAM

NRAM is Nantero's proprietary computer memory technology. It is a non-volatile random access memory based on the location of carbon nanotubes deposited on a chip-like substrate. In theory, the small size of the nanotubes allows for very high-density memory.

Nantero has spent nearly 20 years working on NRAM, which works differently than other memory devices. It is made from layers of carbon nanotubes grown from catalyst particles, most commonly iron. Each NRAM "cell" or transistor is composed of a network of carbon nanotubes and works on the same principle as other non-volatile RAM technologies. Carbon nanotubes that are not in contact with each other exhibit a high resistance state, representing the "off" or "0" state; when the carbon nanotubes are in contact with each other, they exhibit a low resistance state, representing the "on" or "1" state.

Compared with NAND and DRAM, NRAM has lower energy consumption, close to zero power consumption in standby mode, faster writing speed, and unlimited scalability. FRAM cannot break through 100 nanometers, EEPROM is generally more than 60 nanometers, NORFlash is more than ten nanometers, and NRAM can be advanced to 5 nanometers, and there is relatively large room for future expansion.

Another major advantage of NRAM over traditional flash memory is its endurance, which allows for almost unlimited read and write cycles. They are also resistant to heat, cold, electromagnetic interference and radiation. Nantero said they can be stored at 85 degrees Celsius for thousands of years and have been tested at 300 degrees Celsius for 10 years without losing even a single bit of data.

NRAM can be used not only for data storage but also for program storage. This feature is very attractive to the consumer electronics market. Currently, product development projects for standalone NRAM and embedded NRAM are underway. Standalone NRAM is being pursued for three purposes: for DRAM replacement, for NAND flash replacement, and for applications where neither DRAM nor NAND flash is addressable. In the field of embedded memory, there is ongoing work to use embedded NRAM to replace embedded non-volatile memory, including embedded flash memory or embedded RAM (SRAM or DRAM).

In 2016, Fujitsu and USJC announced that they had reached a consensus with Nantero and obtained NRAM technology authorization to carry out the development, design and production of NRAM. As the first generation of NRAM products, Fujitsu's 16Mbit DDR3SPI interface products are expected to be launched around 2021.

Summarize

With the rise of AI, Internet of Things and other fields, the application of big data is becoming more and more widespread, and these new fields have given rise to new demands for storage. Fast reading speed, high storage density, long life, low voltage, and smaller size have become the most urgent needs at present, but the current several types of storage are no longer adequate.

This also provides new opportunities for the above-mentioned five storage technologies. No matter which storage technology, each has its own uniqueness and has huge advantages over flash memory. Among them, MRAM has become the most optimistic technology among semiconductor analysts due to its rich types, broad application prospects, and obvious comprehensive advantages.

But this does not mean that MRAM is a sure winner. With the development and application of other storage technologies, there is the possibility of replacing it. It remains to be seen which storage technology is the future.

access:

Jingdong Mall