Universal Memories Fall Back To Earth

By Mark LaPedus
Ten years ago, Intel Corp. declared that flash memory would stop scaling at 65nm, prompting the need for a new replacement technology.

Thinking the end was near for flash, a number of companies began to develop various next-generation memory types, such as 3D chips, FeRAM, MRAM, phase-change memory (PCM), and ReRAM. Many of these technologies were originally billed as “universal memories.” By definition, a “universal memory” is a single product that could replace all four conventional memory types: DRAM, NAND, NOR and SRAM.

As it turned out, conventional memory has scaled much further than previously thought, pushing out the need for the next-generation technologies. And, in fact, most next-generation memory types are still in R&D. They are expensive to make and difficult to scale.

While there is a frenzy of activity in next-generation memories, the rhetoric surrounding the “universal memory” is fading. Because of the complexity and soaring I/O requirements in today’s systems, there is no single next-generation memory type that has the cost benefits of DRAM, the speed of SRAM, and the non-volatility of flash.

“It is unlikely that we will see a universal memory,” said Gill Lee, a senior director and principal member of the technical staff at Applied Materials. “I do not see a new memory type can replace both NAND and DRAM.”

Clearly, after years of hype, the universal memories have come back down to earth. In possibly the best-case scenario, a next-generation technology could become a mere one-to-one replacement for today’s memory types, Lee said. “We might see them in certain applications and segments,” he said.

Universal niches
In the meantime, HP, Intel, IBM, Micron, SK Hynix, Toshiba, Samsung and others are developing various next-generation memory types. Because it remains unclear which technology will replace DRAM and flash, the larger players are developing most next-generation memory types.

There is a sense of urgency to develop these technologies. DRAM could stop scaling somewhere at the 1nm node. “There is not much room for the floating gate to scale in flash. Nobody really believes that planar NAND can go below 10nm,” Lee added.

In one possible scenario in the next five years (or longer), a next-generation MRAM called spin-torque MRAM (STT-RAM) is the candidate to replace DRAM and SRAM. Also in the distant future, 3D NAND and ReRAM may replace NAND flash and the disk drive.

And PCM could not only displace NOR, but it could also emerge as a new class of so-called storage-class memory. MRAM and ReRAM also are being positioned as storage-class memories, which supposedly fit between the main memory and the processor to alleviate the I/O bottleneck in a system.

Alan Niebel, chief executive of Web-Feet Research, has a different viewpoint. The next-generation memory types are classified as storage-class memories, which can be sub-divided into two groups: memory (DRAM-like) and storage (NAND-like), Niebel said. “Possibly by 2020, one technology may be able to bridge the cost, performance, persistence, and power parameters to satisfy both memory and storage needs,” he said. “In the meantime, the leading replacements for NAND in storage are phase-change and ReRAM. STT-RAM could be a DRAM replacement, but it is too costly for storage.”

NAND and DRAM replacements
If or when planar NAND runs out of gas, the prevailing school of thought is that 3D NAND will replace NAND, followed much later by ReRAM. ReRAM is non-volatile and based on the electronic switching of a resistor element material between two stable resistive states. Startup Adesto is sampling one form of ReRAM, dubbed conductive bridging RAM (CBRAM), which is an EEPROM replacement. HP, Micron, Samsung, SK Hynix and others are working on NAND-replacement ReRAMs.

The first ReRAMs are based on a 1T1R (1 transistor and 1 resistor) structure. Next-generation ReRAMs are based on a 1R structure and consist of various architectures, such as 3D and cross-point arrays. These ReRAMs present several challenges, prompting some to believe that these memories won’t appear until 2015 or so. “Each of the metal layers requires advanced lithography, which is very expensive,” said David Eggleston, senior vice president at Rambus. In 2012, Rambus acquired ReRAM developer Unity Semiconductor.

At a recent event, SK Hynix outlined its strategy, which typifies the roadmap of a NAND vendor. First, SK Hynix will continue to extend planar NAND. “I think scaling NAND to 12nm will be very challenging,” said Sung Wook Park, executive vice president and head of the R&D Center at SK Hynix.

SK Hynix is separately developing a 12nm planar NAND part and 3D NAND. 3D NAND is targeted as the successor to planar NAND, Park said. In addition, the company is also working on next-generation STT-RAM with Toshiba. SK Hynix is separately co-developing PCM with IBM.

The industry also is keeping a close eye on SK Hynix and Hewlett-Packard, which have been jointly working on commercializing HP’s memristor by 2015. A form of ReRAM, memristor is a passive two-terminal electronic device. In memristance, if the flow of a charge is stopped by turning off the applied voltage, this component will “remember” its last resistance.

Initially, devices based on the memristor are aimed for storage, said Janice Nickel, research manager at the Cognitive Research Laboratory at HP Labs. “Then, we will look to move up from there.”

SK Hynix has developed an 8-Mbit test chip based on the memristor. HP itself has demonstrated a 54nm cross-bar structure. “The challenge is the integration of new materials,” Nickel said.

Others hope to ship ReRAMs sooner than later. Micron and Sony, for example, have been co-developing so-called Adaptive ReRAM for possible introduction in 2014. Adaptive ReRAMs are expected to have up to 8-Gbit capacities. Initially, Adaptive ReRAM is geared for cache module applications, said Keiichi Tsutsui, senior manager of advanced memory systems at Sony.

Like NAND, the industry is searching for a DRAM successor. When the DRAM runs out of gas, 3D-based Wide I/0 technology is one possible successor. In addition, Micron and Samsung are developing a 3D DRAM technology called the Hybrid Memory Cube (HMC).

There are several challenges to develop 3D DRAM. Longer term, STT-RAM may replace DRAM. Everspin, IBM-TDK, Qualcomm-TSMC, Samsung, Toshiba and others are working on STT-RAM.

STT-RAM makes use of a spin-transfer torque technology. This is an effect in which the orientation of a magnetic layer in a magnetic tunnel junction (MTJ) can be modified using a spin-polarized current. STT-RAMs are fast and non-volatile, but the challenges include scalability and unstable switching currents in the MTJ memory cell.

“It’s really too early for MRAM to replace DRAM,” said Phillip LoPresti, president and chief executive of Everspin Technologies, an MRAM supplier. “MRAM is always going to be behind in cost and density.”

For now, MRAM is geared for the embedded market. Everspin, for example, is shipping first-generation MRAMs based on a toggle-write technology, mainly for the battery-backed SRAM replacement market. In addition, Everspin is also readying the world’s first STT-RAM. In a slide at a recent event, Everspin called it a ST-RAM or ”SpinRAM.”

Using an alternate method for programming an MTJ element, ST-RAM is mainly geared to replace “battery-backed DRAM” or persistent RAM in hard drives and related storage applications, said Steffen Hellmold, vice president of marketing at Everspin. In an invited paper at the upcoming IEDM, Everspin will describe how they built the largest functional ST-MRAM circuit ever built, a 64-Mbit device with good electrical characteristics.

For main-memory in PCs and other systems, DRAM will remain the dominant technology for some time. “MRAM will not replace DRAM for at least the five years,” Hellmold said. “I am willing to place a bet on it.”

New memory phase
Like MRAM and ReRAM, PCM is in its infancy. PCM is difficult to scale and limited by the power required to change from the crystalline to the amorphous state. Researchers are looking at germanium telluride (GeTe) materials to overcome these limitations, said Jean-Luc Delcarri, general manager of Altatech, a subsidiary of Soitec.

Gary Kotzur, a distinguished engineer at PC maker Dell, said PCM has a potential place in online transaction processing (OLTP) systems. For OLTP, PCM needs to have “faster writes,” he said. “The power must be lower.”

Another application is online analytical processing, but for this application, “we need much higher densities,” he added.