MRAM Begins To Attract Attention

By Mark LaPedus
In the 1980s, there were two separate innovations that changed the landscape in a pair of related fields—nonvolatile memory and storage.

In one effort, Toshiba invented the flash memory, thereby leading to NAND and NOR devices. On another front, physicists discovered the giant magnetoresistance (GMR) effect, a technology that forms the basis of hard disk drives, magnetoresistive random-access memory (MRAM) and other products.

Needless to say, flash memory and disk drives are enormous industries. So what about MRAM? To begin with, MRAM is a promising technology that delivers the speeds of SRAM and the non-volatility of flash with unlimited endurance. At one time, MRAM was even billed as a “universal memory,” which is a single product that could replace today’s DRAM, flash and SRAM.

MRAM, however, is just now getting off the ground after years of delays, and the technology is no longer being promoted as a “universal memory.” In fact, Everspin—the MRAM spinoff of Freescale—is the only vendor that is shipping MRAM chips in volumes, mainly for niche-oriented embedded applications.

But now, several suppliers are gearing up for a long-awaited, second-generation MRAM technology called spin-transfer torque MRAM (STT-MRAM). STT-MRAM is not only targeted for embedded applications, but also storage-class memories (SCMs) and cache memory. In addition, STT-MRAM is also targeted to replace DRAM or flash, but that could take five years or longer before that occurs.

“Potential applications for STT-MRAM are in the areas of MCUs, storage controller buffers, and later on, low-level cache with high-speed perpendicular magnetic anisotropy,” said Subramani Kengeri, vice president of advanced technology architecture at GlobalFoundries. “We can expect the first embedded applications potentially as an embedded flash replacement.”

Integrating MRAM and its offshoot, STT-MRAM, in a system is easier said than done, as there are a number of challenges with the technology. “The higher current toggle switching mechanism in traditional MRAM is limited in terms of density and scaling,” Kengeri said.

What is MRAM?
MRAM uses the magnetism of electron spin to provide non-volatility. STT-MRAM is an effect in which the orientation of a magnetic layer in a magnetic tunnel junction (MTJ) can be modified using a spin-polarized current. In addition, there are several types of STT-MRAM. One type, perpendicular STT-MRAM (p-STT MRAM), utilizes perpendicular magnetization. Another type, in-plane STT-MRAM, involves in-plane magnetization.

“P-STT is targeting embedded flash, and is a potential replacement for SRAM in cache, mostly L3 cache. In-plane STT is primarily a flash replacement,” Kengeri said. “The most scalable version is p-STT, but it is quite complex and the least explored. We think there is a consensus building that in-plane STT-MRAM may be ready for production soon, and p-STT in the next few years. In our view, the former is less scalable for many reasons, including switching stability.”

In a simple MRAM manufacturing flow, a wafer is processed in a standard CMOS step. Then, the MRAM process modules are integrated in the back-of-the-line (BEOL) process. “For MRAM, some of the technology challenges include the integration and manufacturing of non-traditional materials and multilayered stacks,” he said. “The key technology challenges are magnetic film stack deposition, magnetic film etch, and the integration with a standard BEOL.”

The manufacturing challenges, coupled with delays in the technology, have led to a considerable shakeout in the MRAM market. Today, the STT-MRAM players include Avalanche, Crocus, Everspin, Samsung and Spin Transfer Technologies. There are also a number of MRAM alliances that have formed over the years—IBM and TDK; Hynix and Toshiba; Micron and A*STAR; and Qualcomm and TSMC. And recently, GlobalFoundries joined Imec’s STT-MRAM program, which also includes Qualcomm.

MRAM gets off the ground
So far, Everspin is the only vendor to ship MRAM in volumes. Instead of chasing after the “universal memory” market, the company took a pragmatic strategy. Initially, its first-generation, toggle-based MRAMs were targeted for the battery-backed SRAM replacement market. Everspin’s MRAMs have also made inroads as an alternative to nonvolatile RAM for RAID controllers and storage systems.

Everspin’s next challenges are to ramp up the industry’s first STT-MRAM and migrate to 300mm wafers. Rolled out last year, the company’s STT-MRAM has a DDR3 interface and boasts a memory bandwidth of up to 3.2-GBytes/second. “The big market for STT-MRAM is storage,” said Phillip LoPresti, president and chief executive of Everspin. “We see demand for persistent memory.”

Everspin has a 200mm fabrication line in Chandler, Ariz., where it produces MRAM on 180nm and 130nm processes. The company is moving its production on 300mm wafers to undisclosed fabs. “With 300mm, you get more die per wafer,” LoPresti said.

Another STT-MRAM vendor, Crocus, is looking at different applications that may go into production next year. The startup has developed a proprietary 32-bit microcontroller, based on a self-reference thermally assisted switching (TAS) MRAM technology. “It’s a complete MCU,” said Bertrand Cambou, chairman and chief executive of Crocus.

The next wave of applications for STT-MRAM includes three markets—storage-class memories (SCMs); cache; and DRAM and/or flash replacement. In theory, an SCM sits between a processor and main-memory DRAM in a system. The SCM offloads many of the functions of the DRAM. Phase-change, ReRAM and STT-MRAM are billed as SCMs.

Meanwhile, Micron, Samsung, SK Hynix, Toshiba and others are scrambling to develop standalone STT-MRAM devices and for good reason. Some believe DRAM will hit the wall at the 1xnm node, prompting the need for STT-MRAM or another replacement technology.

To date, however, STT-MRAM is behind the curve. “STT-MRAM density is a major issue,” said Alan Niebel, chief executive of Web-Feet Research. For example, Samsung recently moved into production with DDR4 DRAMs, based on 4-Gbit densities and 20nm-class processes. In comparison, the most advanced STT-MRAMs in the market are 64-Mbit parts, based on 65nm or 40nm technology.

Samsung is working on a 1-Gbit STT-MRAM device, which is due out in 2015 or 2016, according to Niebel. To help accelerate its efforts, Samsung recently launched the Global MRAM Innovation Program, which is looking for new breakthroughs in STT-MRAM. The company is accepting proposals from universities and research labs. Selected proposals will receive financial support from Samsung.

On top of that, the industry is taking a hard look at another long-term application for STT-MRAM—cache memories. SRAM is the key component used to make cache memories for microprocessors in PCs and mobile products. SRAM is fast, but it is also expensive and occupies a large amount of board space.