The name is Cerfe Labs (see ARM forms spin-off to pursue CeRAM memory) and one might be tempted to say “get in line” except there is something a bit different about this startup, the claim of bulk-switching between metal and insulator.
Back in 2014 I gave the opinion that ARM’s move into non-volatile memory research was a good thing (see ARM’s turn to non-volatile memory is right move).
Now ARM, the intellectual property licensing parent, has pushed its fledgling NVM research out of the nest, in the form of Cerfe Labs, headquartered in Austin, Texas. In so doing it has hopefully created a nimble startup that may be able to fly and take bulk-switching forward to a significant place in semiconductor development
ARM is a minority shareholder in the startup and is bank rolling it for now, so it could continue to benefit from progress made there, although further dilution of its interest through rounds of venture capital is to be expected. It could be argued that after five years of research ARM is giving up on a potentially significant and fundamental materials technology.
But there again significant brain power is not giving up on the technology. The fledgling includes numerous ARM alumni including former ARM Fellow Greg Yeric; Eric Hennenhoefer, a former vice president of research at ARM; and distinguished engineer Lucian Shifren who has spent several years leading the materials research project at ARM.
Yeric observed in email correspondence with eeNews Europe: “ARM would like to see this technology be successful but they are a few steps removed from the initial users, which is why they are spinning us out.”
This makes some sense but such an exodus of talent from the research pool at ARM does give pause for thought about ARM under SoftBank Group ownership, and uncertainty caused by Nvidia’s offer to buy the company, which may or may not go through (see Opinion: Nvidia’s bad deal is not yet done).
Next: ARM’s loss is Cerfe’s gain
But whatever the background, ARM’s loss is Cerfe Labs’ gain and the technology under development is certainly intriguing.
It is a technology that I and one or two others have been keeping an eye on for the best part of a decade. It relies on the doping of a transition metal oxide that can be “born-on” as a metal but when hit with an applied voltage and critical current density switches to an insulating state. This so-called correlated electron or electron orbital Mott transition is reversible, inherently fast and takes place throughout the bulk of a material. It should also have a low current density during the setting process.
That in turn brings likely benefits over the myriad forms of ReRAM that are currently out in the wild and attracting investment. But research proofs grind exceedingly slowly and researchers have limited access to the tools needed to make critical dimensions in line with leading-edge manufacturing processes.
So far the research team has moved the original Symetrix Corp. (Colorado Springs. Colo) work at 5 microns diameter down to 47nm and have developed more than four transition metal oxide material systems. What Cerfe needs to do over the next couple of years is to prove these switches at 30nm, 20nm and 10nm. By then leading foundry TSMC is likely to be offering a nominal 3nm manufacturing process with a 10nm half-pitch.
Greg Yeric, CTO at Cerfe Labs, is bullish about the technology’s ability to scale. He said: “Deep dimensional scaling coupled with multi-level cell capability and 3D crossbar stacking capability are combined, without trade-off, with fast and low power switching, and an unheard-of temperature window. All that comes in a one-film deposition that we have shown works in spin-on, PVD [plasma vapor deposition] or ALD [atomic layer deposition]. So a large swath of memory applications are on the table at this point, from edge AI to automotive to HPC [high-performance computing], and they can flexibly be addressed in manufacture from spin-on with possibly etch free process to high precision ALD.”
Yeric offered that the team’s nanometer-scale devices show a greater than 50x on/off ratio and are created in the low-resistance state with forming-free operation. His assessment is that the indicators are that the 2nm CMOS compatible bit cell with a select transistor is possible, with multi-level cell operation as a bonus.
Next: Stacking up
But the CeRAM could also go to multi-layer stacking, Yeric indicated.
“Regarding stacked arrays, we think it looks attractive compared to existing stacked embodiments (PCM, possibly ReRAM), but we think a non-hafnium based FeFET would ultimately be the best solution for cost, as 3D NAND has demonstrated.” This point of view gives a reason why Cerfe Labs has also been given a license to develop ferroelectric FET devices by research partner Symetric Corp. (Colorado Springs, Colo.).
“Our focus is on CeRAM for the combination of scalability, speed, and power, especially in the embedded memory space. It is showing the potential to be ‘more universal’ than other options, which is why we are excited about it.” But Yeric added that a third terminal would be needed for vertical stacking. Yeric said that while CeRAM could potentially be made into a FET it may be more straightforward for FeFET. “FeFET may not be quite as fast as CeRAM or scale to the same densities, but there is a lot of space for it to contribute to edge AI,” Yeric said.
But in one important aspect of this potentially important memory element we have less information than we did five years ago and earlier. And that is of what it is likely to be made.
Going back to 2011 the early research was on carbonyl-doped nickel oxide as a potential bulk-switching RRAM non-volatile memory. Now Cerfe is saying it does not use nickel-carbonyl as a pre-cursor material and it has more than four CeRAM materials of proven capability and – for now – Cerfe isn’t revealing those materials and their respective dopants and electrode materials.
Others have speculated that they may include hafnium-oxide, titanium-oxide and yttrium-titanium oxide besides nickel-oxide.
Yeric was prepared to talk about the high degree of controllability and flexibility Cerfe thas in the making of such memories.
Devices are now manufacturable using spin-on slurries, plasma vapor deposition (PVD) or high-precision atomic layer deposition (ALD), he said. “You can go after ultralow-cost IoT with the slurries, and you can go after high density memories with MLC [multi-level cell] using ALD. A sweet spot for most embedded applications may be the in-between, PVD.”
A manufacturer could also vary the film thickness to tune the current and vary the doping chemistry and doping concentration. This could also be used to manage other trade-offs such as the switching voltage and voltage margins.
The technology also shows remarkably low temperature variability. This means that devices can work at close to absolute zero and at high temperatures. Testing to date has demonstrated performance from 1.5K to 115 degrees C, Yeric said. There is an expectation that the CeRAM switch will operate at up 400 degrees C.
This allows CeRAM to be a contributor to some cryogenic quantum computing control systems. “There is about a 50 percent fall off of current from 100K toward 1.5K but even at 1.5K it is a very good memory device and we expect to show millikelvin capability. This stability appears to apply to both voltage and current,” Yeric said.
It may therefore be that CeRAM memories tuned for either extremely low temperature or extremely high temperature could be some of the first to appear on the market.
Next: Partners and ecosystem
Wonder materials for non-volatile memories have been pitched for decades. They range from the 40-year R&D gestation of chalcogenide phase change memory through carbon nanotubes and on to the umpteen flavors of ionic migration and filamentary resistive RAM we see today.
As such technologies get closer to commercial viability their developers are sometimes reluctant to expose details by academic publication. As a result the first sign of success for such technologies tends to be the sign-up of partners and customers who have done private due dilligence. This may be particularly true in the case of Cerfe, which needs partners for pilot-scale and/or commercial-scale wafer processing to show what it can do. So the names of the first strategic partners, investors and customers are awaited with interest.
As an aspiring intellectual property licensor Cerfe Labs is not so different its parent. But to judge from previous memory technologies it has long road ahead. If all goes well it might be able to contribute in a major way to the 2nm semiconductor node for which a number of materials decisions are up for debate.
Cerfe needs active, enthusiastic partners to prevent it ending up as a curiosity at the side of the semiconductor roadmap. But as the company is staffed with ARM alumni the building of an ecosystem in support of their technology should be second nature.
Related links and articles: