Healthy markets are possible for emerging memories without trying to supplant DRAM or NAND flash...
It is time for a frank discussion about emerging memories. Many have now been percolating for decades with the promise of displacing established incumbents such as DRAM and NAND flash.
But will emerging memories ever see the light of day? Despite research breakthroughs and new patents on potentially disruptive technologies, DRAM and NAND technologies continue to advance despite the slowing of Moore’s Law. That means the goal posts are always moving for the possible replacements. MRAM, ReRAM, FRAM and PCRAM are often discussed in the context of emerging use cases such as automotive, industrial Internet of things (IoT), edge computing and sensor nodes, even AI and machine learning. But they often fall short where its counts: reliability and longevity.
There are in fact niche applications for next-generation memories, and the opportunities are not dissimilar to legacy technologies that still reap healthy profit margins in smaller market segments. The reason? Legacy systems often represent the best solution for a specific problem.
Emerging memories have been around for decades. While some have found a measure of commercial success as embedded technologies, they have also lagged as cost-effective alternatives to discrete memories. That, despite higher performance, endurance and retention, or reduced power consumption.
Magneto-resistive random-access memory (MRAM) was first developed in the 1980s and promoted as a universal memory. Unlike other memory technologies, MRAM stores data as magnetic elements rather than electric charge or current flows. Performance-wise, MRAM is similar to SRAM because of its use of sufficient write current. But that dependence also hampers its ability to compete at higher densities with DRAM and flash.
While MRAM pioneers such as Everspin have seen some success in the embedded market for discrete applications, and even demonstrated it can handle the extreme environments of automotive applications, MRAM remains a niche memory.
Similarly, resistive random-access memory (ReRAM) has yet to mature as a viable discrete memory; even its success in the embedded market has been limited. Adesto Technologies, recently acquired by Dialog Semiconductor, was one of the first companies to bring commercial ReRAM devices to market with its CBRAM technology. Its appeal included lower power consumption, fewer processing steps and lower voltages when compared to conventional embedded flash technologies. It also exhibited radiation tolerance for space and medical applications.
Several companies have been developing ReRAM technologies over the last two decades, but the approach still faces integration and reliability challenges. Like magneto-resistive, ReRAM vendors have made some inroads in developing embedded ReRAM devices to drive revenue that can fund discrete development efforts. In collaboration with research partner Leti, Weebit Nano announced in late 2019 it would ramp efforts to solve the selector problem necessary to make discrete ReRAM commercially viable. Meanwhile, it continued to explore the memory’s potential for neuromorphic and AI applications. CEO Coby Hanoch has previously told EE Times Weebit is still a memory start-up that must build a revenue stream from embedded products in order to advance on other fronts.
The appeal of its ReRAM technology is that it leverages materials that can be used in existing production lines.
The other emerging memory seeking to flourish as a discrete technology is ferroelectric RAM (FRAM), which uses a ferroelectric instead of a dielectric layer to achieve non-volatility. While fabrication steps are similar to DRAM, FRAM functionality is more like flash.
FRAM is arguably the most successful of the emerging memories in that it has made headway in embedded applications and has the potential to reach higher densities. Around for about 35 years, FRAM’s nonvolatility and low power consumption are desirable characteristics for many applications, and there are small niche applications.
For example, Cypress-Infineon offers its Excelon FRAM for automotive and industrial applications, with densities as a high as 8 Mb in low-pin-count, small-package options. The Excelon family was specifically designed for the high-speed, nonvolatile data logging needed for autonomous vehicles. It is also used in medical, wearable, IoT sensor, industrial and other advanced automotive applications. Significantly lower power consumption, data retention and radiation resistance make FRAM a viable replacement for EEPROM and NOR flash. Implanted medical devices that must function for up to a decade is one example.
Elsewhere, Ferroelectric Memory Company (FMC) of Germany foresees higher density, viable storage-class memories. FMC is exploring the potential of hafnium oxide to help create the larger transistors needed for the technology to achieve higher densities that can be cost-effectively manufactured.
The ability to control costs while scaling manufacturing is critical if emerging memories are to become viable alternatives to DRAM and flash, even if it’s only for niche applications. None of the appealing characteristics of MRAM, ReRAM, FRAM or PCRAM ultimately matter if manufacturing costs are excessive. Even 3D NAND went through growing pains — despite offering numerous benefits over its planar predecessor.
Phase-change memory (PCRAM) in the form of 3D Xpoint jointly developed by Intel and Micron might have finally reach a tipping point that allows it to graduate from the “emerging” category. PCRAM has a long history, but only in the last decade has all that research moved into commercialization. PCRAM is non-volatile and exploits the unique behaviour of chalcogenide glass. So far, its one real commercial success has been in Intel’s Optane. Like 3D NAND, PCRAM has had its growing pains.
However, 3D Xpoint has proven more cost-effective to manufacture. Moreover, chip makers other Intel and Micron might be able to build products around it. Intermolecular Inc., a subsidiary of Germany’s Merck, recently disclosed research around a new combination of materials that could enable a 3D vertical non-volatile memory architecture, shaping the second iteration of 3D Xpoint technology.
Intel currently offers Optane as both DIMMs and SSDs, and has positioned it as storage-class memory. The company seems comfortable with the technology replacing DRAM and flash in some use cases, but not all. For example, the Optane Persistent Memory 200 Series comes in the form of DIMMs aimed at replacing NAND SSDs for real-time analytics applications. The reason is speed: 225 times faster than a mainstream SSD.
Christopher Tobias, general manager of Intel’s Optane Solutions Division, said one advantage of SSDs was their ability to handle random input/outputs. That was at a time when CPUs had just a few cores and virtualization were just getting started. With multiple cores, virtualization and software containers taking off in the cloud, the challenge has become feeding all these cores.
As DRAM approaches scaling limits, Tobias said, there is an opportunity for Optane to augment DRAM, serving CPUs in a DIMM format. “We’ve been producing it at scale, and it serves a critical need,” the Intel manager said. “We can augment memory in feeding all these virtual cores and compute becomes denser.”
Optane has reached the point where it’s cost-effectively filling a layer between NAND SSDs and DRAM, said Tobias, whether its complementing DRAM or meeting latency demands that traditional SSD cannot.
“The big cloud companies have a real need to fill in these layers.” Hurdles remain, Tobias added, since software and architectures need to catch up to fully take advantage of Optane’s performance gains.
“Bringing a new silicon technology to market is very daunting. There’s multiple challenges to overcome,” like scaling manufacturing. The challenges of transitioning a working prototype from the lab to the production line, he stressed, are “enormous.”
Deep pockets and experienced engineers are required. So too are long-term commitments and patient capital, according to Michael Yang, senior director for memory at research firm Omdia. Few companies other than Intel have been able to push an emerging memory technology across the finish line. Optane “can turn a corner because there is a market,” Yang said.
The market is not replacing DRAM or flash, but using alternative memory where it makes sense. Yang said there’s a lesson for other emerging memory makers in Intel’s Optane approach: finding a profitable market for technologies rather than trying to displace a proven memory. “For [technologies aspiring] to replace DRAM or flash, that train left the station years ago.”
Most MRAM companies have been around too long to be considered start-ups, making little progress beyond embedded markets — let alone supplanting DRAM or flash.
Claiming you can replace DRAM or flash might make for great marketing pitches, but it is not realistic from an engineering perspective. Most companies simply can’t be the John Wick of memory, said Yang, but there is business potential in niche memory markets such as embedded and low-density devices. Hence, it may be time to jettison the “emerging” label, whether it’s MRAM, ReRAM or FRAM. “They’re just little memory,” Yang said.
Thomas Coughlin has been tracking the emerging memory market for the past three years in collaboration with Objective Analysis. Coughlin said embedded is the best opportunity for many of the emerging memories. Even if it’s a standalone, a device is going to be low density for applications that require low power consumption. “There’s a lot of applications that would benefit enormously from a little bit of non-volatile memory,” he said.
DRAM and NOR flash may eventually hit scaling limitations. That would create an opening in embedded applications for FRAM, MRAM, and ReRAM once they pass qualification screening.
Whether the moniker is still apt or not, MRAM and ReRAM are still emerging memories, but Coughlin said Optane has emerged since Intel has likely reached a break-even point for Optane manufacturing. Meanwhile, its customers are realizing the need for fast, inexpensive memory, like DRAM, he said. “It’s an inexpensive memory expansion capability. It allows lower cost, memory-intensive applications.”
While Optane may be a billion-dollar product given its suitability for workloads like real-time analytics, said Yang, other memory makers need to come to grips with the fact that their market potential is smaller, but healthy.
“It’s not black or white,” he added. “You can be successful at $30 million. It’s all different shades of success.”