Memory Hierarchy Shakeup

Gaps in the memory hierarchy have created openings for new types of memory, and there is no shortage of possibilities.


It’s no secret that today’s memory chips and storage devices are struggling to keep up with the growing demands in data processing. To solve the problem, chipmakers have been working on several next-generation memory types. But most technologies have been delayed or fallen short of their promises.

But after numerous delays, a new wave of next-generation, nonvolatile memories are finally here. One technology, 3D NAND, is shipping and gaining steam. And three others—Magnetoresistive RAM, ReRAM and even carbon nanotube RAMs—are suddenly in the mix.

These next-generation technologies could alter the memory landscape. Over time, the new memory types could augment or replace the traditional memories, such as DRAM, NAND and SRAM. The new memory types also will impact the existing memory/storage hierarchy. Today’s hierarchy is fairly straightforward. SRAM is integrated into the processor for cache. DRAM is used for main memory. And disk drives and NAND-based solid-state storage drives (SSDs) are used for storage.

There’s a problem, though. There are gaps in the hierarchy. “We see the continued explosion of data. (There is a) need to get access to that data to turn that into information to solve very real problems. That is about getting access to that data quickly in large amounts and in a timely manner,” said Rob Crooke, senior vice president and general manager of Intel’s Non-Volatile Memory Solutions Group, at a recent event.

“This explosion of data says that solid-state drives based on NAND technology coming from the storage side of things is not enough,” Crooke said. “And we’ve been working for quite some time on what comes after that and what comes in addition to NAND-based technology.”

What’s next is a technology called 3D XPoint, at least according to Intel and its development partner, Micron. Considered a resistive random-access memory (ReRAM), 3D XPoint could augment or replace today’s NAND and DRAM.

Still, there are big questions about 3D XPoint and the other next-generation nonvolatile memory types. What will these technologies do? And how do they fit in the hierarchy?

There are no simple answers. To help OEMs get ahead of the curve, Semiconductor Engineering has provided a description of some of the new memory types and the challenges ahead.

Today’s planar NAND is the mainstream memory technology for flash drives and SSDs. The latest NAND chips are at the 1xnm node. But at 1xnm, planar NAND is running out of steam, prompting the need for 3D NAND.

3D NAND resembles a skyscraper, in which horizontal levels are stacked and then connected using tiny vertical channels. So far, the big market for 3D NAND is SSDs.

Samsung, the leader in 3D NAND, is shipping its third-generation 3D NAND device. The device is a 48-layer, 3-bit multi-level-cell (MLC) technology, resulting in a 256-gigabit chip.

Samsung’s previous device was a 32-layer chip. “From a performance standpoint, (the 48-layer device) is roughly 2X to 2.2X on a sequential read and a sequential write,” said Jim Elliott, corporate vice president of Samsung Semiconductor. “It basically uses less than half the power as its predecessor.”

Meanwhile, the other suppliers, Micron/Intel, SanDisk/Toshiba and SK Hynix, are sampling their initial 3D NAND devices. Most of these vendors are expected to move into production later this year. “Memory investments focused primarily on 3D NAND capacity additions (are) expected to increase and broaden out among the market leaders in the second half of the year,” said Rick Wallace, president and chief executive of KLA-Tencor, in a recent conference call.

Still, 3D NAND is taking longer than expected to become a mainstream technology. “Progress is indeed being made, but it’s not moving faster toward the mainstream. Samsung is forcing itself down the learning curve by producing parts earlier than other manufacturers would. The others are now sampling, but they are still a long way from volume production,” said Jim Handy, an analyst with Objective Analysis.

“When manufacturers were saying they would produce in volume in 2016, I was guessing that it would be delayed until 2017. Now that they are starting to call for volume in 2017, I am projecting 2018,” Handy said. “I am saying 2018 because I anticipate cost-effective production to be considerably more difficult than anyone anticipates.”

To scale 2D NAND, chipmakers require leading-edge lithography. In contrast, 3D NAND relies on an assortment of complex deposition and etch steps.

The 3D NAND flow starts with a substrate. Then, thin films are stacked layer by layer on the substrate using alternating stack deposition. This process is much like making a layer cake. “The key challenge is making sure that each pair of layers is exactly the same as the one above it,” said Rick Gottscho, executive vice president of global products at Lam Research.

Then, high-aspect ratio trenches are etched from the top of the device to the substrate. The aspect ratios for 3D NAND are 40:1 to 60:1, compared to 12:1 or 15:1 for planar NAND. “In the case of NAND, you have this huge inflection because of 3D NAND, where the aspect ratios are going through the roof,” said Amulya Athayde, senior director of global product management at Applied Materials.

After the trenches are formed, the device requires contacts. The device is backfilled with a conductor like tungsten using a metal deposition step. “That’s a tricky deposition, because you are doing a non-line-of-sight deposition,” said Dave Hemker, senior vice president and chief technology officer at Lam Research.

Still, the future looks bright for 3D NAND, at least in the near term. But long term, there are fears that 3D NAND may hit the wall and stop scaling. “3D NAND will hit physical scaling limits after 128 layers,” said Alan Niebel, president of Web-Feet Research. “Some believe it could scale infinitely. This is unrealistic and wishful thinking.”

For years, ReRAM has been touted as the replacement for flash and other memory types. ReRAM is nonvolatile and based on the electronic switching of a resistor element material between two stable resistive states. The technology is attractive because it delivers fast write times with more endurance than today’s flash.

ReRAM falls under three categories—embedded, 2D and 3D. Within the 3D category, there are two types of ReRAM—cross-point and vertical.

Today, Panasonic and Adesto are shipping ReRAMs for embedded applications. The 2D- and 3D-based ReRAMs could one day replace flash, but these technologies are difficult to make and are late to the market.

But according to the announcements from Intel and Micron, ReRAM-like 3D XPoint could appear sooner than later. Based on a cross-point architecture, 3D XPoint is up to 1,000 times faster and has up to 1,000 times greater endurance than NAND, and is 10 times denser than conventional memory. Available in 2016, Intel will incorporate a 3D XPoint device in an enterprise SSD and a DDR4 DIMM for servers.

“Initially, 3D XPoint has RAM-like characteristics at a price lower than DRAM,” said Greg Wong, an analyst with Forward Insights. “It can potentially compete with 3D NAND in enterprise SSD applications, where low latency and high IOPs are required.”

Others agree, but there is still some doubt whether 3D XPoint is really a pure ReRAM. “XPoint may not be ReRAM, but a new storage-class memory with features from phase-change memory, ReRAM and others,” Web-Feet’s Niebel said.

“First, XPoint will be used in SSDs and then DIMMs. And later, XPoint will move into mobile for something like UFS or eMMC,” Niebel said. “XPoint will compete with 3D NAND at the high-performance areas, where NAND is trying to do caching and fast-access/high-throughput tasks. XPoint will cannibalize both DRAM and NAND in these performance avenues in the next couple of years.”

Unlike 3D NAND, ReRAM will require advanced lithography, which could get expensive. The other challenges for ReRAM include material complexities, leakage, parasitic capacitance and line resistance.

MRAM uses the magnetism of electron spin to provide non-volatility. It delivers the speed of SRAM and the non-volatility of flash with unlimited endurance.

The next big thing is a second-generation MRAM technology called spin-transfer torque MRAM (STT-MRAM). STT-MRAM 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.

MRAM has been touted as a replacement for DRAM and flash. So far, MRAM hasn’t lived up to those promises, but it is making inroads in the embedded market.

“A lot of people ask us how does MRAM compete with the other technologies,” said Phillip LoPresti, president and chief executive of Everspin, a supplier of MRAMs. “We don’t compete with NAND. In a lot of cases, we make NAND better. (ReRAM) can’t do our function. We aren’t trying to do their function. They can all work with each other.”

But will STT-MRAM ever replace DRAM? “DRAM has a fundamental problem. You have to refresh it. People want persistence. DRAM isn’t persistent. DRAM will reach a limit and will only start servicing areas where persistence doesn’t matter. That will probably go on forever,” LoPresti said. “Besides its scalable limits, DRAM will be used less. It will get replaced by other technologies in different areas.”

Today, Everspin, Micron, Samsung, SK Hynix, Toshiba and others are developing STT-MRAM. But to date, Everspin is the only vendor shipping MRAM.

Everspin has been shipping a first-generation MRAM technology called field switch. Field-switch MRAMs are targeted for the battery-backed SRAM replacement market.

In 2012, Everspin started sampling the world’s first STT-MRAM, a 64-megabit device with a DDR-3 interface. Everspin, which refers to its technology as ST-MRAM, has made inroads in the enterprise SSD market, particularly for the write-buffer part of the system.

In general, STT-MRAM faces some challenges, namely cost and scalability. For example, Everspin has a 200mm line in Arizona, where it produces MRAM on 180nm and 130nm processes.

To help solve the challenges, Everspin last year announced a foundry deal with GlobalFoundries. Under the plan, GlobalFoundries will make Everspin’s ST-MRAMs based on 40nm and 28nm processes using 300mm wafers.

With this technology, Everspin will develop higher-density, lower-cost ST-MRAMs, such as 256-megabit and 1-gigabit parts. “Our partnership with Everspin will help drive ST-MRAM adoption and feed the rapidly growing MRAM market,” said Gregg Bartlett, senior vice president of product management at GlobalFoundries.

There are other promising technologies on the horizon. For example, Nantero is ramping up what it calls NRAM. Based on carbon nanotubes, NRAM is as fast as DRAM and is nonvolatile like flash.

Not all technologies will move in the memory/storage hierarchy, however. For example, FeRAM remains stuck in the embedded market. Another technology, phase-change memory (PCM), appears to be fading from the picture.

MRAM, NRAM and ReRAM are more promising. But these technologies could meet the same fate as PCM if they fail to hit the market window at a reasonable cost.