MRAM's manufacturing challenges are unique relative to other emerging memories.
TORONTO — Of all the perennially emerging memories, MRAM seems best poised to be on the brink of significant, broad adoption. Whether or not this will happen soon is dependent both on advances in manufacturing and an ecosystem to support the technology for discrete and embedded MRAM devices.
Objective Analysis and Coughlin Associates’ annual report, Emerging Memories Ramp Up, suggests MRAM, along with PCRAM and ReRAM, has reached a critical point where it makes sense in more applications than ever before. It does, however, have its own set of manufacturing challenges, said co-author Thomas Coughlin, both from a process and materials perspective, as it uses some materials and processes that differ from conventional CMOS manufacturing.
“Right now, MRAM is made as a [back end of line] (BEOL) process in a separate fab. New equipment not used in conventional CMOS manufacturing is needed, such as ion beam etching and new sputtering targets.” For costs to be reduced for embedded MRAM products, he said, manufacturing will need to move into the CMOS fab and become part of the regular device manufacturing.
Beyond incorporating MRAM further into the manufacturing chain, quality control and yield improvement, as with other semiconductor manufacturing processes, will be an ongoing challenge and opportunity, said Coughlin, and the fact that all the large semiconductor foundries have offered MRAM memory as an option for embedded products is significant. “By incorporating MRAM into their embedded products and bringing MRAM into the regular device manufacturing in high volume, yield and quality issues will be resolved, and the unique tools used for MRAM production will become more common place and more embedded into foundry production. This will lower costs and increase availability.
One of those manufacturers, Applied Materials, recently announced updates to its platform, specifically addressing the challenges that are unique to MRAM, including the need for novel materials. It has evolved its Endura platform from a single process system to an integrated process system as part of its materials engineering foundation for emerging memories, including MRAM.
Applied Materials has evolved its Endura platform to address the materials deposition challenges that come with MRAM manufacturing (Source: Applied Materials)
Kevin Moraes, vice president of metal deposition products for Applied Materials, said the biggest manufacturing challenge for MRAM has to do with the complexity of the stack and the number of layers needed — in excess of 30. “The reason it’s pretty complex is because these layers serve multiple purposes.” Fundamentally, MRAM is essentially comprised of little narrow magnets, he said, so magnetic materials are required that can preserve a certain orientation, including a bottom reference layer, that’s immune to any external magnetic fields.
There are also material aspects with that stack and several layers that are used either as barrier layers or seed layers, said Moraes, and then there’s the extremely thin MgO layer that’s intrinsic to making the tunnel junction, which is at the heart of the MRAM stack. “But because this barrier is very thin, there’s also a risk that it can break down easily,” he said. “It demands a lot of perfection. It demands a lot of capability of many layers, many materials. You need to have that precision so that you can deposit exactly the right thickness.”
Jeff Lewis, Spin Memory’s senior vice president of business development, said Applied Materials has hit the mark with Endura as the quality of deposition is critical to the performance of the MRAM device itself, and in general, the foundries have created the tool sets that are enabling MRAM to go into manufacturing. He said this helps to create an environment where companies can start designing specifically for an MRAM device for artificial intelligence and Internet of things (IoT) applications where its persistence, power gating, and power management are critical. “MRAM has some unique capabilities there.”
Its capabilities can lead to MRAM replacing existing memories, such as SRAM, or creating new use cases all together, said Lewis. Spin Memory is looking at making MRAM as SRAM-like as possible, in that it offers the same value proposition but with three to four times more memory in the same footprint, without any of the leak that comes with SRAM. But as much as continually refining materials is important, Spin Memory is taking a circuit-level approach that works with any magnetic tunnel junction, he said, and that’s enabling it to make orders of magnitude improvements in terms of endurance, which in turn improves performance.
Everspin has two types of MRAM in production.
STT MRAM requires controller enablement or an FPGA,
and the company is continuing to grow its ecosystem
through partnerships. (Source: Everspin)
Everspin, meanwhile, has a history of collaboration with GlobalFoundries, and it’s been able to fine-tune its manufacturing for more than a decade in its own fab by constantly tweaking the recipe, said Troy Winslow, vice president of global sales for Everspin. “It has given us the ability to quickly ramp and transfer that knowledge to GlobalFoundries.” The company has two types of MRAM in production — toggle MRAM and STT MRAM, the latter of which does require controller enablement or an FPGA.
“We’re continuing to grow our ecosystem to support the STT marketplace, and provide simple and quick time-to-market implementation and adoption of our spin torque in the marketplace.”
At the 2019 Flash Memory Summit, Everspin announced several partnerships: Phison Electronics and Sage Microelectronics will provide native support for its 1 Gb STT-MRAM memory, while Cadence Design Systems will provide DDR4 design IP and verification IP (VIP) support for Everspin’s 1 Gb STT-MRAM. These collaborations will enable system designers to use STT-MRAM in their products with their choice of using standard controllers, designing with FPGAs, or building their own SoC/ASIC.
An ecosystem that supports MRAM will help reduces its costs, said Coughlin. “Having third-party controllers capable of managing devices that contain MRAM will help more companies participate in using MRAM in their systems and thus increase the overall yield and demand for MRAM and reduce the costs of MRAM memory.” He sees the biggest potential volume of MRAM in embedded applications. “Moving MRAM to embedded applications will force foundries to move MRAM manufacturing into the conventional CMOS device manufacturing. This has challenges but also provides huge opportunities in the cost of MRAM and its uses.”
As much as price affects which memory is adopted, Winslow said MRAM doesn’t necessarily have to be cheaper than incumbent options if a slight premium price leads to a lower overall total cost of ownership. He sees MRAM following the path of SSD adoption in that they were considerably more expensive than a hard drive but were able to create a new market because they had capabilities that had value in certain applications. “Our customers are seeing there is a value above and beyond the bit cost. It is a total cost of ownership story. They can articulate and justify paying the MRAM premium over existing technology.”