MRAM, where to go?

Magnetoresistive RAM (MRAM) appears to be gaining traction at the most advanced nodes, partly due to recent improvements in the memory itself and partly because new markets require solutions for which MRAM may be the only qualified solution.

There are still many skeptics about MRAM, and there are many potential competitors. Over the past few decades, MRAM's role has been limited due to high cost, low density, and lower durability. But the number of supporters is growing.

Figure 1: Annual shipments of independent memory from 2022 to 2033

(in petabytes)

Today, there are two main types of MRAM: standalone and embedded, and three main application areas:

• Non-volatile RAM (NVRAM), which is where eMRAM competes with embedded SRAM and embedded DRAM;

• eMRAM, which can be designed on the most advanced nodes to compete with flash memory, and

• Framebuffer MRAM, usually targeted at monitors.

“Frame buffer MRAM is used in photography or video and is tuned for durability and speed,” said Martin Keim, senior engineering director for Tessent at Siemens EDA. "The trade-off is that the volatility is down to just a few seconds. But if you just take a picture, you need to store it inside the silicon for a second and then move it to permanent storage. Ultimately, the advantages of different MRAMs are actually depends on what you want to build.”

Foundries offer different MRAM variants, each with different PPA metrics. “The market space requires different eMRAM variants to provide certain PPA metrics,” said Bhavana Chaurasia, product manager at Synopsys. “For example, one might be more like SRAM, with higher speed and smaller area, but with shorter retention periods, rather than Volatile eMRAM has higher retention, is slower and has a larger area, so now you have the flexibility to decide which type of eMRAM to use for your SoC.”

MRAM is not a perfect solution on leading-edge nodes, and it certainly won't replace DRAM (which was our original goal). "Several emerging memory technologies have advanced to the point where they can produce Mb to Gb densities," explained Gary Bronner, senior vice president at Rambus Labs. "These include phase-change memory such as PCM or 3D Xpoint, MRAM and ReRAM. However, in order to replace DRAM, these memories need to have similar performance to DRAM, but cost more. There is currently no emerging memory that can prove the cost and performance required to replace DRAM. MRAM and ReRAM are looking for alternatives to embedded flash memory. is a higher performance, higher cost option, and ReRAM is a cost-effective alternative.”

Still, MRAM’s footprint appears to be growing. New developments help overcome its historical limitations. At the same time, flash memory is limited on leading-edge nodes, while MRAM is particularly suitable for automotive applications.

"MRAM has become very important to the microelectronics industry because of the lack of solutions to replace flash memory at advanced logic nodes," said Sebastien Couet, IMEC's ​​Magnetics Program Director. Most major foundries have essentially been in production for the past two to three years. STT MRAM. This is an important first step because the introduction of new memory technologies in the industry is very rare. SRAM, DRAM and flash have come a long way over the past 50 years, but fundamentally they are the same technology. We reached high-volume manufacturing capabilities [for MRAM], which is a very important milestone. It is expected to grow significantly in the next 5 to 10 years. "

As a magnetic technology, MRAM is inherently radiation resistant. This makes stand-alone versions popular for aerospace applications that are also less price sensitive. "It's relatively large, and in the memory world, size means cost," Couet said. "That's good for the spacecraft, which can buy a few megabytes of memory for €5,000, but obviously it's not good for the consumer market."

MRAM has also found a place in enterprise storage, such as IBM's Flash Core Module (Figure 3, column 1), where Everspin's MRAM is used as a buffer in the event of an unexpected power outage.

In addition, MRAM is also used in industrial applications, said Jim Handy, principal analyst at Objective Analysis. As an example, he cited how well-designed industrial robots can be prevented from colliding with each other when restarting after a power outage by maintaining correct arm positions.

“Some of these applications need to be very fast to write and they need to be non-volatile,” Handy explains. "Today's high-capacity non-volatile memories are NAND flash, NOR flash and EPROM, which is a very small market. All three are very slow to write and consume a lot of energy, which makes them useless for some needs. Not popular for fast data storage applications. Another option is SRAM, which comes with battery support, but this is not popular because the battery needs to be replaced every few years.”

But the automotive industry has really tipped the balance, moving from flash to eMRAM as demand for MCUs increases and the cost of creating flash memory, which can require a dozen mask adders, increases.

"Below the 32nm node, the cost of co-integrating flash with the CPU rises dramatically because the manufacturing methods and logic are different," imec's Couvet said. "You need to have specific steps in the fab to just make flash memory. It does get crazy to the point where flash manufacturing costs more than other manufacturing costs, so the industry is looking for solutions."

The solution is eMRAM. In 2022, Renesas Electronics solved this problem when it announced the launch of STT-MRAM test chips at the 2002 VLSI Symposium, “Compared with flash memory made with FEOL, MRAM made with BEOL has advantages in processes below 22 nm , because it is compatible with existing CMOS logic process technology and requires fewer additional mask layers," the company said at the time.

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 will no longer be embedded flash in advanced nodes, all generations Factories have stopped developing it. The transition period is 22nm and 28nm, depending on the foundry."

Embedded MRAM is also inherently scalable. "eFlash has reached the scale limit below 28 nanometers, and with eMRAM you can design embedded non-volatile memory that supports lower technology nodes," Chaurasia said. "The benefit of this is scalability of core and memory flexibility. The memory can be kept as part of the same chip, which helps achieve smaller area and better performance and power consumption compared to other options, whereas keeping the memory in a separate chip results in the Performance and power consumption loss and increased safety issues eMRAM has smaller area, lower leakage, higher capacity and better radiation resistance. Compared with PCRAM and ReRAM, eMRAM has lower temperature sensitivity. Provides better production-grade yields and provides longer endurance (data retention across multiple read/write cycles over many years). It allows word-level erase and program operations, making it an energy-efficient NVM solution. . All these advantages make eMRAM an excellent eNVM.”

All of these advantages could make eMRAM the future of automotive memory, or so its proponents believe. However, not everyone agrees.

"When we talk to automotive customers, they prefer ReRAM to MRAM," Keim said. “The arguments we hear are about temperature stability and concerns about magnetic fields in the automotive environment affecting the data stored in MRAM. For the former, today’s MRAM types do have the temperature stability (-40C to +150C) required by the automotive market. This shows that MRAM technology is rapidly expanding into all corners of the application field. How far their productization has progressed so that automotive users can choose products off the market is a different question."

MRAM, on the other hand, has a very narrow gap between the legal read values ​​of 0 and 1 (see tuning discussion below). This small gap narrows and changes with increasing temperature, making it more difficult to safely interpret the readings. The resistance gap between 0 and 1 is much larger in ReRAM, making it easier to read, i.e. easier to operate at higher temperatures.

Even so, in May this year, NXP and TSMC announced the launch of the first 16nm finFET eMRAM for automobiles, which can update 20MB of code in about 3 seconds, compared to 1 minute for flash memory, providing up to 1 million update cycles , endurance levels that are 10 times higher than flash memory and other emerging storage technologies.

The competition between old and new memory types is more than just a technical and financial one. Imagine if the ambulance had to pause for a minute before it could start.

Like many devices in the tech industry, MRAM started with Nobel Prize-winning work. In this case, it was the discovery of giant magnetoresistance, the concept underlying hard drive storage. The next breakthrough was spin transfer torque (STT), an effect that can use spin-polarized currents to modify the orientation of the magnetic layers in magnetic tunnel junctions (MTJs), or spin valves.

This new physics leads to more efficient MRAM. The MTJ at the core of STT MRAM consists of two magnetic electrodes with a dielectric tunnel barrier between them. "Basically, the MJT is responsible for detecting the magnetic state of one of two electrodes," Couet explains. "In MRAM, you have a fixed magnet, then a tunnel barrier, and then a free magnet that you can program to point up or down. And then you compare. After that, the resistance of the junction depends on the arrangement of the electrodes. . If the two electrodes are aligned in a parallel configuration with respect to their magnetic state, they have low resistance. If they are antiparallel, they have high resistance. This provides a way to differentiate between magnets facing up and down. You can store information this way and the magnets are stable for long periods of time.

STT MRAM currently dominates. Field-switched MRAM (now called switched MRAM) is an early form, first brought to market in 2006 by Motorola Freescale, which was later spun off and merged into Everspin Technologies. Everspin offers most of the standalone MRAM on the market today in switched and STT forms.

"In toggle MRAM, the polarity is changed by applying a magnetic field to the MTJ," explains Joe O'Hare, Everspin's senior director of marketing. "In contrast, STT MRAM technology uses spin-transfer torque properties, or manipulation of electron spins through polarizing currents."

The current interest was sparked in 2010 when Worledge and his colleagues at IBM, as well as the Ohno Laboratory at Tohoku University in Japan, published papers demonstrating vertical STT MRAM. “Until then, MRAM was just a laboratory curiosity,” Worledge said. "You can make some bits, but they don't switch reliably. The reason is that all the magnetization is in-plane, but that's very inefficient. What you want is the magnetization perpendicular to the plane of the wafer."

Although the answer has been known in theory since 1996, it took more than a decade to craft the right combination of physics and materials for optimal switching. "The vertical bit solves the writing problem," Warrich explained. "That's what's causing all this excitement. All the big companies are jumping on board and starting to work on STT RAM."

There is also a new variant called spin-orbit torque (SOT)-MRAM, which imec has been improving. As Couet explains in the blog, "The main difference between STT-MRAM and SOT-MRAM is the current injection geometry used for the writing process. In STT-MRAM, the current is injected vertically into the MTJ, while in SOT-MRAM The current injection occurs in the plane of adjacent SOT layers.”

Still, the complex physics mean designers need to be aware of the complexities of using MRAM.

"When you buy SRAM or DRAM off the market, they are more or less the same," Keim said. “MRAM has different underlying physics and properties, and if you choose the wrong ones, you can run into trouble.”

MRAMs have probabilistic behavior, so they also have a lot of error correction capabilities. "Typically, your ECC might have one bit for error correction and two bits for error detection. MRAM goes a step further and has at least two bits for correction and three bits for detection," Keim said. "Then ECC will correct for the impossibility of MRAM. So you get deterministic behavior, just like SRAM, but it comes at a cost to you."

There is an additional question. "When you turn on the SRAM and write and read, it just works. With MRAM, the read and write circuits first have to be trained on the actual properties of the cell, which is called fine-tuning," he said. "Trimming means Then you have to find your comparison value, your resistor value, which lets you decide whether what you just read should be interpreted as a 0 or a 1. Therefore, MRAM requires a complete trimming cycle to read and write. "

It's not over yet. "Different parts of the memory require different tuning values. It's a complete process to make this thing really work," he said. "Once you've done that, you have enhanced ECC, you've done the tuning circuitry that needs to be on the chip. There's so much data that needs to be processed off-chip, which is time-consuming and expensive. So you have to invest in additional hardware for simple learning. The only feasible solution is to do all this on-chip, fully automated and independent. To further reduce area costs, you need to leverage an existing DFT, the Memory Built-in Self-Test (MBiST) engine. "

The good news is that after all the extra time and money invested, the memory works pretty well.

Heat dissipation is the weak point of DRAM. Higher temperatures can cause its capacitors to refresh and lose data. "DRAM is a difficult technology to work with. It forgets things in milliseconds and requires periodic offline refreshes," said Marc Greenberg, group director of product marketing at Cadence. "It's susceptible to heat and radiation."

In contrast, one of the big reasons for eMRAM's success is that it remains stable at high temperatures to the point where data can be retained through reflow soldering. However, its thermal advantage also explains its relative lack of durability.

“Because you need to have very good data retention to withstand reflow, you have to make the magnetic bit difficult to switch because you don’t want it to switch during thermal fluctuations,” IBM’s Worledge said. "This means it is also difficult to switch when writing, so a larger voltage must be applied to write. Over time, larger voltages can damage the magnesium oxide tunnel barrier."

Every technology has its Achilles' heel. For MRAM, it is a strong magnetic field.

“As with any magnetic component, eMRAM should be kept at a distance from other magnetic components,” Chaurasia said. “For example, there may be inductive coils within the SoC or chip. A certain distance between these devices reduces the magnetic effect. Therefore, it must be maintained while designing the chip. For off-chip magnetic immunity, advanced packaging can provide shielding. eMRAM blocks Or an air gap above the device can also provide protection from magnetic susceptibility."

Unlike being near an MRI machine, which is an absolute no-no, proximity to small motors and small induction coils shouldn't be a problem, IBM's Worledge said. Still, he did raise a nightmare scenario - an attacker using NdFeB magnets could disable eMRAM chips, so it's best to place such chips at least a centimeter from the surface of a device, especially a car. .

There are many recent and long-term developments in MRAM's work. According to IBM's Worledge, mobile cache may become available within the next year or two, with MRAM replacing flash and SRAM. "From an R&D perspective, I think we're done. We've demonstrated a magnetic tunnel junction that meets these requirements, and we hope the foundry will have a commercial product available soon."

Worledge believes there will be a killer application in the future of eMRAM. "Let's say you have an IoT device that's powered by solar energy. Maybe it's part of a security system. It's usually off, and maybe it has a little sensor on it that picks up some audio. When the audio gets loud enough, it wakes up And try to use some kind of artificial intelligence algorithm to determine if there is an intruder. Without eMRAM, it will wake up, load the weights of the AI ​​program from flash memory into SRAM, and then run the AI ​​program if the weights are updated. Back to flash memory, which is very power-hungry. But with just eMRAM, when the device wakes up, it starts running because all the weight is already in the eMRAM."

This approach is slower than SRAM, he said, but has the advantage of lower power consumption compared to writing data back and forth between flash and SRAM, and its durability makes it ideal for field operations.

The holy grail that IBM has been working on for nearly a quarter of a century is last-level caching. Current caching schemes rely on SRAM as working memory and can suffer from a problem called "cache miss", where the data is not in SRAM. This forces SRAM all the way to DRAM to get the data.

"It's a very slow process, about 35 nanoseconds or 50 nanoseconds round trip," Worledge said. "You do want to have a larger SRAM cache if you can. If MRAM is twice as dense as SRAM, it will have twice as many bits in it. It will be a little slower than SRAM, but it will have twice as many bits in it. Twice as fast as SRAM, which more than makes up for the slower speed.”

Warrich believed they were finally within striking distance. "This is a very challenging application because you need MRAM to be very dense, very fast and very durable," he said. "We still need to reduce the switching current. We published a new device we invented, called a dual-spin-moment magnetic tunnel junction, which indeed reduced the switching current by a factor of two. We even demonstrated 250 picosecond switching, with It's incredibly fast compared to all other MRAM publications, but there's a lot of promise there, but we're still in the research phase."

*Disclaimer: This article is original by the author. The content of the article is the personal opinion of the author. The reprinting by Semiconductor Industry Watch is only to convey a different point of view. It does not mean that Semiconductor Industry Watch agrees or supports the view. If you have any objections, please contact Semiconductor Industry Watch.

Today is the 3490th issue shared by "Semiconductor Industry Observation" with you. Welcome to pay attention.