TORONTO – Resistive RAM is the epitome of an “emerging” memory in that, for the most part, it’s still in the research and development phase.

“I’d say the most progress is being in R&D,” said Jim Handy, principal analyst with Objective Analysis, and co-author of the report Emerging Memories Ramp Up. There are many small companies working to commercialize it, such as Crossbar and Weebit Nano, and there’s also a lot of research work being done at LETI, a technology research institute of CEA Tech based in France. The DRAM manufacturers are also dabbling in it, he said, but no one is rushing to bring out a part, although every foundry that offers MRAM also has some sort of ReRAM capability.

Essentially, the major memory makers are hedging their bets. “They’re making sure that whenever DRAM and NAND flash all run out of steam that they’ve already got some research effort going on to supersede it,” said Handy. That leaves smaller companies with the task of funding the R&D to see if they can hit a tipping point that will lead to large scale commercial applications. Aside from Adesto Technologies’ trademarked CBRAM, which the company was founded on, he said, there’s very little ReRAM in production.

ReRAM’s ability to help emulate how the brain works to create neural networks for artificial intelligence applications is fueling a lot of the interest in the emerging memory. “Neural maps are actually something that people have been talking about for a very long time,” said Handy. “But nobody really has implemented something and put it into production.” However, there’s also talk that neural networks could be based on standard flash, he said. “They could abandon ReRAM, if flash can do it.”

Crossbar is a founding member of a consortium recently formed to create AI platforms using ReRAM. (Source: Crossbar.)

At IEDM 2019, Leti presented research work outlining how it has fabricated a fully integrated bio-inspired neural network, combining ReRAM-based synapses and analog spiking neurons, while measuring a 5x reduction in energy use compared to an equivalent chip using formal coding. The neural network implementation is made such that synapses are placed close to neurons, which enables direct synaptic current integration.

Earlier this year, ReRAM maker Crossbar Inc., along with others, formed SCAiLE (SCalable AI for Learning at the Edge), an AI consortium dedicated to delivering an accelerated, power-saving AI platform. The group is focused on combining ReRAM with advanced acceleration hardware and optimized neural networks to create ready-made, power-efficient solutions with unsupervised learning and event recognition capability.

Outside of the neural network hype, said Handy, there are two ways to look at the ReRAM market; one is standalone parts, like an EERAM, where it’s self-contained in a chip, and the other is as an embedded memory inside of SoCs, perhaps in a microcontroller. The standalone applications are appealing as a replacement for the SRAM with backup battery combination so that information isn’t lost during a power failure, he said, while embedded scenarios are more interesting because NOR flash has hit a wall at 28 nanometers. “Because of that the major foundries and anyone who makes a micro-controller are looking at what they can do to allow the memory those things to shrink beyond 28 nanometers. Both ReRAM and MRAM offer a lot of promise there.”

Weebit Nano Ltd. is one company that’s keen on using ReRAM for neural networks, but even CEO Coby Hanoch said the more mundane applications will likely hit the market first. Regardless, he sees ReRAM really moving forward in general. “Some of the bigger companies are now putting more emphasis in it. And then you have the startups that are making progress.” Weebit specifically is confident enough about its technology that its talking to customers. The company recently announced it’s working with XTX Technology, Chinese provider of memory solutions for consumer electronics, industrial embedded systems, telecom, and networking applications to incorporate Weebit’s silicon oxide ReRAM technology. “We’re heading towards productization and commercialization. I won’t be surprised if by 2021 we’ll start seeing some of the first real commercial deals and moving ReRAM forward.”

Weebit’s ReRAM cell consists of two metal layers with a Silicon Oxide (SiOx) layer between them comprised of materials that can be used in existing production lines. (Source: Weebit Nano.)

Hanoch said the increased interest in general for productizing ReRAM has been a critical milestone in the past twelve months, and for Weebit, it was shoring up its technology by achieving more than a million cycles of endurance. “We now have retention 10 years at 150 degrees,” he said. “So, the technology is becoming robust and it’s really ready to go out.”

While Weebit has been quite public about its efforts on the neural networking front, Hanoch said the company’s first phase is the embedded market. “ReRAM has significant advantages in the embedded space.” That includes replacing flash technology and external non-volatile memories with a faster, lower-power ReRAM option that he said is better than MRAM solutions because Weebit’s silicon oxide ReRAM uses standard materials that easily fit into existing production lines. “We’ve seen that we can manufacture it with just one or two added masks.”

Earlier this year, Applied Materials announced its Endura Impulse PVD platform that enables precise deposition and control of the multi- component materials used in ReRAM. (Source: Applied Materials.)

Manufacturing a memory chip is also a critical factor in its success. For ReRAM, uniformity and interface control are both critical in the manufacturing process, said Mahendra Pakala, managing director, Memory Group, Advanced Process Technology Development at Applied Materials. Earlier this year, the company announced its Endura Impulse PVD platform for ReRAM includes up to nine process chambers integrated under vacuum along with on-board metrology to allow the precise deposition and control of the multi-component materials used in ReRAM, which are essential to realizing high performance, reliability, and endurance in high volume, he said. “There are at least five, six layers in the conductive chambers with different materials and housing, and different processing conditions so they can form the right interfaces.”

Pakala said right now a key focus for ReRAM is looking at different material systems as it gets bigger. “It’s still early enough in this stage of development that we are looking at material screening now.” Like Weebit, Applied Materials sees ReRAM getting traction in embedded applications while there’s more work to be done to make discreet ReRAM products commercially viable. To help build the ecosystem necessary to ramp up ReRAM and other emerging memories, he said. The company opened its Materials Engineering Technology Accelerator (META Center) to facilitate collaboration between researchers, equipment suppliers, and manufacturers at the earliest stages of development — essentially to speed up the “lab to fab” process.

Whether it’s embedded or discrete, a big part of ReRAM’s appeal is the lower power consumption — the recent Leti announcement is an example — which makes it ideal for IoT applications, said Weebit’s Hanoch. “ReRAM has some advantages, in terms of speed and lower power, that are compelling.” The embedded space represents the lowest hanging fruit for ReRAM, he said, so that’s where the company’s focus is in the short term, but it also has “aggressive” plans for discrete products too.

Hanoch said as more commercial deals are realized and customers gain confidence, ReRAM will hit the bend in the hockey stick toward fast growth, so to speak, in the next three to four years. “That’s when it’s really going to happen.” However, the neuromorphic applications that are getting a lot of buzz are still at least five years out, he said. “There’s still quite a bit of research to be done before it can be productized, but that will probably be the biggest market for ReRAM once it’s established. The potential is just unbelievable.”