New memory finds a lucrative niche between other existing memory types as competition grows.
Resistive RAM appears to be gaining traction. Once considered a universal memory candidate—a replacement for DRAM, flash and SRAM—ReRAM is carving out a niche between DRAM and storage-class memory. Now the question is how large that niche ultimately becomes and whether other competing technologies rush into that space.
ReRAM (known alternately as RRAM), is a type of non-volatile memory that began garnering attention in 2009 when startup Unity Semiconductor emerged from stealth mode. Rambus bought Unity in 2012 because it was one of several contenders for the next generation of memory technology, along with ferroelectric RAM (FeRAM) and Magnetoresistive RAM. ReRAM also has been considered a possible replacement for 2D NAND, NOR flash, and other memory types.
Since then, multiple competitors have entered into the ReRAM business, which seems to validate the potential here.
“I have worked in memory all my career, and for years it was looked down upon as boring,” said Gary Bronner, vice president of Rambus Labs. “Today it is leading innovation. It’s very exciting.”
He’s not alone in that view. “It’s a really exciting time to be in the memory business,” said Sylvian Dubois, vice president of strategic marketing at Crossbar.
What makes ReRAM so interesting is the limitation of other memory choices. There is DRAM for rapid access memory; NAND flash, which is three orders of magnitude slower; and there is storage-class memory in between. Storage-class memory, a term first coined by IBM several years ago, could have a huge impact on computation efficiency.
“I believe that the gap could be filled by two or three different types of ReRam, and could see a real reduction in the volume of DRAM being used,” Bronner said. “This would have very significant impact on the industry. Architects have been very clever at taking advantage of developments. They already take advantage of hierarchy of memory on chip and chip to chip.”
There is a decent list of alternative memory types that all rely upon a bi-stable material as the storage element that changes resistance. Rambus is working with a multilayer metal oxide structure that changes resistance by injecting ions into the material.
Crossbar uses silver atoms suspended in an amorphous silicon matrix. Under write voltage, atoms from a top silver layer migrate into the matrix to form bridges of conductive metal filaments. “These filaments are only 3nm in diameter, but create a very large on/off ratio,” Dubois said. The company has published results for a 7nm read cell.
The other option that has been widely publicized involves phase-change materials, which depend on melting a material and then cooling, quickly or slowly, to create either crystalline or amorphous phases. In terms of these materials choices, Bronner observed that “the physics of the phase change material is probably the best understood.”
However, thermal-based solutions have had a rocky history in the semiconductor industry, from Heat Assisted Magnetic Recording (HAMR) to smectic liquid crystal displays. The problem is that heat spreads, so there is crosstalk between neighbors and DC heating of the part that depends on duty cycle.
“Many people have favored other materials over phase change for these reasons,” Bronner said. “However, Intel and Micron have pioneered work in these materials, presumably because the device physics is much better understood.”
In the end heat may limit the phase change device to lower bandwidths and higher power, according to Dubois.
In addition to material choices, there also are architectural choices. Although crossbars have received a lot of publicity, a 1 transistor-1 storage cell similar to DRAM is where many are starting, particularly for embedded memory. It makes integration much simpler and gives the best access times.
“You can add the new material after the conventional device processing is completed,” Bronner said.
Crossbars give higher density and single bit access, but are limited in the size of each block of crossbars. This is a similar problem to the old multiplexed displays, where the cell count depends on the non-linearity of the on-off switch.
The third choice is a 3D stack. One possibility is multilayers of crossbars, but this requires lithography at each layer. More interesting, in Bronner’s view, is “an equivalent to 3D NAND, where one litho step creates a vertical string of storage cells. 3D NAND will allow flash to continue to scale for five to seven years, and then new materials will have an opportunity.”
The penalty for storage in the form of strings is that there is no longer single-bit access, so the memory slows down. Different access times and cost structure could result from single bit, byte, and multi-byte architectures.
Dubois emphasized that crossbars and 3D all require a multiplexed 1 transistor to N cell architecture, which in turn means that each cell must have a non-linear element. Some teams use a diode with each cell, but Crossbar has demonstrated a cell that has its own non-linearity.
Who and what will win?
“The most impressive progress was disclosed by the Intel/Micron partnership over the summer when they described a 128Gbit chip 3D XPoint” memory,” said Bronner. “To even think about building a device of this size, requires a very mature level of process and defect control.”
Intel had a previous false start with phase-change memory. The company made it clear that the technology has shifted, but has not elaborated on that, according to industry sources.
A search of the U.S. Patent and Trade Office patent application database found crossbar memory materials patents applications as recently as 2012 assigned to Micron that focused on metal chalcogenide phase change systems, which suggests that 3D XPoint may well be an improved phase change system.
In its announcement Intel claimed “a crossbar structure which is 1,000 times faster than NAND flash and 10 times the density of DRAM.” The company also showed a patterned wafer, discussed an operational manufacturing plant in Utah, and said it plans to sell product next year.
Panasonic currently sells a tantalum oxide-based ReRam embedded flash replacement for on-chip static memory.
Rambus purchased the ReRam IP of Unity Semiconductor in 2012 for$33M, and has licensed that IPto a number of parties. Unity had raised more than $22M and created 145 patents.
“Rambus is also partnering with licensees, such as Tezzaron, to create embedded flash products,” Bronner said. “The focus of licensing is architecture and materials.”
A patent search suggests those materials are metal oxides.
Elsewhere, in the startup world, Crossbar announced on Sept. 14 that it has completed a $35 million Series D funding round, bringing the total investment so far to $85 million. Crossbar plans to use the funds to continue the commercial ramp of its “game-changing non-volatile (NVM) memory technology.” At IEDM in 2014, the company reported a device architecture that “has been successfully demonstrated in a 4 Mbit integrated 3D stackable passive crossbar array.”
Dubois said Crossbar received wafers from one of its production manufacturing partners. “The new funding will allow us to put products with embedded RAM on the shelves and move Crossbar forward.”
It appears that differences between the various competitors are primarily in their storage material choices that determine power, access time, read/write cycles and cell size. This is a competition that requires deep pockets, and smaller players are relying on being able to use existing fabs to compete with the industry giants who can bankroll a custom fab.