TORONTO – A flavor of resistive RAM (ReRAM) has overcome a significant hurdle that has it nipping at DRAM's heels.
4DS Memory Limited recently announced that architectural changes to its patented Interface Switching ReRAM have improved read access so dramatically that it is now comparable to DRAM. In a telephone interview with EE Times, company CEO and Managing Director Guido Arnout said the development places the company in a hallway with a lot of doors it could potentially walk through.
The challenge for most emerging memory technologies, including ReRAM, has been inherently high bit error rates, which in most cases, he said, is caused by large random cell current fluctuations. Error correction techniques have traditionally been used to reliably retrieve the data, but these are typically time consuming and negatively affect read access time and cripple read speed.
Arnout said that an extensive current fluctuation study has not found any large current fluctuations with 4DS' Interface Switching ReRAM, which means it requires minimal error correction. This enables high-density storage class memory with effective read speeds comparable to DRAM that previously weren't attainable. Over the past year, he said the company has also achieved scalability to 40nm and significantly boosted endurance.
The “holy grail" was to create a storage class memory that could compete with NAND flash, said Arnout, but that technology has dropped in price quickly. Rather, the opportunity for 4DS' ReRAM is between DRAM and flash. “The space between DRAM and storage is so huge," he said.
4DS' ReRAM includes a patented cell structure and operation called MOHJO — a Metal Oxide Hetero Junction — and a non-filamentary switching mechanism.
There are number of different approaches to developing ReRAM. 4DS's Interface Switching ReRAM differs from filamentary ReRAM in that its cell currents scale with geometry, said Arnout. Smaller cells yield lower cell currents, and lower currents can flow reliably through narrow on-chip wires. Small cells and narrow on-chip wires are required to achieve high-density memories. However, he said, the lower cell currents also translate into longer latency, so the cell currents of small cell sizes need to match the required latency to achieve DRAM latency in a high density storage class memory.
After extensive measurements and analysis of its ReRAM cells fabricated in different sizes, 4DS concluded that cell currents needed to be boosted by an order of magnitude to reach DRAM latency at the cell size targeted by the industry for storage class memory, said Arnout.
However, a short cell latency is still insufficient, he said. The latency that matters is the sum of the inherent memory latency and the time required to detect and correct any read errors, through error correction circuitry (ECC) necessary to reliably detect the state of ReRAM cells over time. Arnout said other ReRAM, including filamentary and CBRAM, for example, have struggled with the high bit error rates. Filamentary ReRAM technologies essentially switch at a single point making them susceptible to irregularities in the switching region, so even small changes will result in very large current changes.
4DS' Interface Switching technology, however, reduces the influence of random irregularities within the switching region, thereby minimizing current fluctuations. Essentially, said Arnout, latency of the company's Interface Switching ReRAM is dominated by the inherent memory latency and not by the time overhead of error correction.
In the memory hierarchy as outlined by research firm Objective Analysis, ReRAM technologies such as 4DS' are looking to fill the space between DRAM and NAND.
It's been a decade since SanDisk predicted that ReRAM would replace NAND flash, but 3D NAND is close to meeting its yield and cost targets to make it competitive with planar flash, and string stacking would appear to provide it with several generations after 64-layer. Arnout said there remains a vast opportunity between DRAM and NAND; it doesn't make sense to see ReRAM has a complete replacement for an existing memory. “There's one rule in the industry: if you create memory that works, it will find a sweet spot. It's about slicing up the space so you can do more things," Arnout said.
He sees an opportunity for ReRAM to provide a tier of storage class memory to enable data centers to deliver new content on the Internet more quickly and efficiently, similar to how flash was a premium tier for more frequently accessed data in the early days of hybrid storage arrays. “There' no end to the imagination when you have this kind of technology," he said.
Now that 4DS has proven the concept of its Interface Switching ReRAM, the focus for the foreseeable future is to scale it, said Arnout, and achieve decent yields.
Jim Handy, principal with Objective Analysis, said the proof of the pudding is always in the eating when it comes to emerging memory technologies. “Nothing is a true contender until it ships in volume," Handy said. "A lot of these technologies have shown promise." He said 4DS has in fact solved a hurdle known to impede the performance of ReRAM, but we're still not meeting the potential long ago predicted for ReRAM.
It has had some niche successes, noted Handy. Adesto Technologies' CBRAM is an ReRAM that works well in applications where radiation tolerance is required, such as medical and space applications, as DRAM and flash aren't options. Carbon nanotubes also fall under the category of ReRAM. “All of these technologies are great for certain niches," Handy said.
Many people would argue that phase change memory (PCM) is ReRAM, he added, which has led to speculation that Micron and Intel's joint 3D XPoint is in fact ReRAM. Handy said Intel has always dabbled in ReRAM, shipping products at a loss because it can make up the difference with processors, something other companies can't do, while 4DS would like to emulate what 3D XPoint is looking to achieve. Handy speculates Intel has set up conditions so that if for any reason a different ReRAM, be it 4DS' or someone else's, becomes more feasible than 3D XPoint, it can quickly switch technologies.
“Objective Analysis doesn't look at 3D XPoint as battle that has been won," Handy said.
—Gary Hilson is a general contributing editor with a focus on memory and flash technologies for EE Times.