Next-Gen Memory Ramping Up

The next-generation memory market is heating up as vendors ramp a number of new technologies, but there are some challenges in bringing these products into the mainstream.

For years, the industry has been working on a variety of memory technologies, including carbon nanotube RAM, FRAM, MRAM, phase-change memory and ReRAM. Some are shipping, while others are in R&D. Each memory type is different and is targeted for specific applications, but they all promise to displace one or more of the conventional memories in the memory/storage hierarchy in today’s systems.

In the first tier of this hierarchy, SRAM is integrated into the processor to enable fast data access. DRAM, the next tier in the hierarchy, is used for main memory. Disk drives and NAND-based solid-state storage drives (SSDs) are used for storage.

Fig. 1: Memory hierarchy—DRAM/SRAM and flash have opposing characteristics that leave a gap to be filled by storage-class memory. Source: Lam Research

Today’s memory and storage types work, but they are struggling to keep up with the explosion of data and bandwidth requirements in systems. For example, DRAM is fast but power-hungry, while both NAND and hard drives are cheap and slow.

That’s where next-generation memory fits in. The new memory types combine the speed of SRAM and the non-volatility of flash with unlimited endurance. These technologies boast some impressive specs, but they have been delayed or fallen short of their promises—or both.

In fact, it has been a struggle to bring many new memory types into mass production. These memories rely on exotic materials and switching mechanisms, making them difficult to fabricate and/or operate in the field. They are also expensive.

All told, the new memories are still niche products, but there is some noticeable progress in the arena. For example, Intel is ramping up a next-generation memory called 3D XPoint. Then, GlobalFoundries, Samsung, TSMC and UMC are developing the new memory types for the embedded market. “The thing that is really a big deal is that logic fabs are developing MRAM and resistance RAM for embedded memory. That’s the thing that might drive the cost out of it,” said Jim Handy, an analyst with Objective Analysis. “For the standalone memory market, the cost is high. And that’s making it appealing only to people who are willing to pay a lot of money, because they can’t get what they want elsewhere.”

So the conventional memories remain the mainstream products in systems, but the new memory types present some options. To help readers get ahead of the curve, Semiconductor Engineering has taken a look at the status of the next-generation memories.

The rise of 3D XPoint
It has taken a long time, but several next-generation memories are ramping up. Each new memory has some intriguing attributes, claiming they can outperform the traditional memories.

Still, it’s unlikely that the new memories will displace DRAM, flash or SRAM, at least for now.

It all boils down to performance, density and cost. For example, the cell size for a given memory equals the feature (F) size times four square. The smallest cell size is 4F². The latest 3D NAND devices incorporate four bits per cell (QLC), which in theory translates into a cell size of 1F².

“If you want to displace NAND, you have to be cheaper than 1F². As far as I know, we won’t see that in our lifetime,” said Ed Doller, a memory veteran and board member at Nantero, a developer of carbon nanotube RAMs.

To displace DRAM, a new memory type must be cheaper as well as having an entire infrastructure around it, such as a DRAM-compatible interface and a controller.

So if the new memory types won’t displace the conventional technologies, where do they fit? “Applications like cloud computing and the latest mobile products are driving the need for a new memory category that combines DRAM’s speed with NAND’s higher bit density and lower cost,” said Alex Yoon, senior technical director at Lam Research, in a blog. “To meet these criteria, several new technologies are being explored. Some are targeting embedded applications such as systems-on-chips (SoCs), while others are focused on the storage-class memory space.”

Currently, the new memory types have carved out niches somewhere in the memory/storage hierarchy that are not met by today’s memories. Some are even taking small bites or share away from DRAM and flash. But it’s unclear if the new memory types ever will become mainstream technologies.

There still is no one new memory technology that can meet all requirements, so customers will likely use one or more over time. “They all have their own place,” said Sylvain Dubois, vice president of marketing and business development at Crossbar, a ReRAM supplier. “In some cases, they are semi-competitive. There is some overlap. But it’s clear that we all have different positioning.”

Fig. 2: Memory hierarchy. Source: Imec

One technology is gaining steam, though. The big change in the market is the recent rise of 3D XPoint, a next-generation technology developed by Intel and Micron.

When 3D XPoint was officially introduced in 2015, the technology was touted as a storage-class memory that fits somewhere between DRAM and NAND. It was supposed to be 1,000 times faster and up to 1,000 times greater endurance than NAND.

In reality, though, 3D XPoint has been delayed and failed to live up to those specs. “We know much of that was over-hyped in reality,” said Mark Webb, the principal at MKW Ventures, a consulting firm. “The reality is different, but it’s still pretty amazing. 3D XPoint will have more revenue than all other new nonvolatile memories combined.”

Indeed, after several delays, Intel is ramping up SSDs and other products based on 3D XPoint. Eventually, Intel will use the technology in a DIMM for servers. “(With 3D XPoint), Intel has the fastest SSDs with the best endurance,” Webb said. “The DIMMS are delayed.”

Sales for 3D XPoint are expected to jump from zero not long ago to $750 million in 2018, according to MKW. By 2020, 3D XPoint revenues are expected to reach $1.5 billion, according to MKW.

In comparison, MRAM was a $36 million business in 2017, according to a report from Coughlin Associates and Objective Analysis. Other memory types are barely on the radar.

The new memories pale in comparison to DRAM and NAND. The DRAM market is expected to reach $101.6 billion in 2018, while NAND will hit $62.6 billion, according to IC Insights.

3D XPoint, meanwhile, is based on a technology called phase-change memory (PCM). PCM stores information in the amorphous and crystalline phases. It can be reversibly switched with an external voltage.

Built around a two-layer stacked architecture, a 3D XPoint device comes in 128-gigabit densities using 20nm geometries. The read latency is about 125ns with 200K cycles of endurance, according to MKW.

Fig. 3: 3D XPoint architecture Source: Wikipedia

The technology is fast, but it isn’t 1,000 times faster than NAND. “It also has a much higher cost than NAND,” Webb said. “It’s not a DRAM replacement. It complements DRAM.”

What’s next for 3D XPoint? The big opportunity is in the DIMM space. When the DIMMs finally appear from Intel, they will integrate both 3D XPoint and DRAM. “Intel can optimize the processor and architecture to take advantage of the performance characteristics of 3D XPoint,” he said.

Still, the future of the technology remains uncertain. Intel and Micron are parting ways in the development of both 3D NAND and 3D XPoint. As previously announced, the two companies will finish the current products in both categories, then pursue these technologies independently.

It’s unclear if Micron will ever field a 3D XPoint device. So far, Micron has yet to ship a 3D XPoint product, as the technology appears to compete with its DRAM and NAND products.

Clearly, Intel has the resources to go it alone in 3D XPoint. But the question is whether Intel will ever recoup its massive R&D investments with the technology.

MRAM vs ReRAM
Meanwhile, the industry is also developing other new memory types, such as MRAM and ReRAM. Like 3D XPoint, MRAM and ReRAM can be made and sold as standalone devices.

3D XPoint is not sold as an embedded memory. In contrast, MRAM and ReRAM can also serve in the embedded memory markets.

For MRAM, the industry is developing a next-generation technology called spin-transfer torque MRAM (STT-MRAM). STT-MRAM uses the magnetism of electron spin to provide non-volatile properties in chips. It combines the speed of SRAM and the non-volatility of flash with unlimited endurance.

Fig. 4: STT-MRAM memory cell. Source: MRAM-Info

In traditional memory, the data is stored as an electric charge. In contrast, MRAM uses a magnetic tunnel junction (MTJ) memory cell for the storage element.

The MTJ consists of a memory stack, which can be re-configured for a given application. But when tuning the MTJ stack, there are some trade-offs in endurance, data retention and write pulse width. “In the design of the MTJ stack, there are inherent trade-offs. You can optimize the stack for endurance by giving up data retention, for example, and vice versa,” said Tom Andre, vice president of engineering at Everspin.

“This allows you to approach different applications in different ways. For example, if you are doing embedded MRAM, and you are trying to build an embedded NVM for code storage, the ability to raise the data retention and give up endurance fits well within this application,” Andre said. “For our standalone parts at Everspin, the trade-off is to go in the other direction. We work more towards higher endurance applications for these circular write-buffer type solutions.”

STT-MRAM has other advantages as well as some disadvantages. “MRAM is great for speed, but it can’t compete on cost or density,” MKW’s Webb said. Plus, STT-MRAM is more difficult to fabricate than previously thought, thereby limiting the shipments of the technology in the market.

To date, Everspin is the only company shipping standalone parts based on STT-MRAM. Everspin is shipping a 256-megabit part based on 40nm, with a 28nm 1-gigabit device in the works. Avalanche, Crocus, Samsung, Toshiba, SK Hynix, Spin Transfer and others are still working on STT-MRAM, but they are not in production.

“I would say the learning curve is very big when you compare magnetics versus semiconductor processes. The learning curve takes time. That’s the biggest problem,” said Mahendra Pakala, managing director of process development at Applied Materials. “For the embedded market, there is enough critical mass. I would say this a lot further along.”

Indeed, the momentum is building for embedded MRAM. GlobalFoundries, Samsung, TSMC and UMC are developing embedded MRAM for foundry customers at 28nm/22nm.

In the embedded market, the industry makes use of microcontrollers (MCUs). MCUs integrate several components on the same chip, such as a CPU, SRAM, embedded memory and peripherals. Embedded memory, such as NOR flash, is used for code storage.

MCUs with embedded NOR flash based on 40nm and above are shipping. Now, the industry is ramping up 28nm MCUs, with 16nm/14nm chips in R&D.

The problem is that it’s difficult to scale embedded flash, sometimes called eFlash, at 28nm and beyond. “(Many believe) that 28nm/22nm will be the end of eFlash, not because of scalability limitations but because of economic barriers,” said David Hideo Uriu, product marketing director at UMC. “Can you scale embedded flash beyond 28nm? The short answer is yes, as we will support it in our 22nm node. But the macro design is essentially the same as our 28nm.

“Beyond 28nm/22nm, eFlash will require over 15 mask adders during front-end-of-line processes. The additional mask adders create economic barriers that challenge the foundry industry whether to pursue alternative nonvolatile memory or to continue investing additional resources to push the boundary of existing eFlash technology,” he said.

So, embedded MRAM is being developed to replace embedded NOR flash at 28nm and beyond. “(Embedded MRAM is) low power with fast write. The write speed is faster compared to much slower NOR embedded flash,” said Mike Mendicino, vice president of leading-edge CMOS at GlobalFoundries.

In one example, a low-power MCU may require fast wake-up and security capabilities. “This is where MRAM could replace the typical embedded flash, but also some SRAM,” Mendicino said.

For cache, SRAM takes up a large part of a chip. Embedded MRAM could also assume some of the SRAM-based cache functions, thereby saving space and cost. “MRAM itself can offer power savings in these devices. But if you put a really good MRAM onto a mediocre platform, that’s not going to win,” he said.

Still, there are some challenges with embedded MRAM, namely the ability to integrate the technology in designs. Cost is another factor. “Customers want the cost of emerging embedded nonvolatile memory to be as cost-effective as eFlash. This expectation presents another challenge to the foundry industry and will be difficult to achieve, but the solutions should be able to maintain current levels of price competitiveness using today’s cost points,” UMC’s Uriu said.

Resistive RAM (ReRAM), meanwhile, also is making progress, but not to the level of 3D XPoint and MRAM. Generally, there are two types of ReRAMs—oxygen-vacancy ReRAM and CBRAM.

In both cases, a switching medium is situated between a top and a bottom electrode. When a positive voltage is applied on the top electrode, a conductive filament forms between the two electrodes. The filament consists of ion atoms. When a negative voltage is applied on the bottom electrode, the conductive filament breaks.

Fig. 5: ReRAM in action. Source: Adesto

ReRAM involves a complicated process. “Understanding that and how to control it is still very difficult and a big challenge,” Applied’s Pakala said.

Both MRAM and ReRAM have similar read and data retention specs. But MRAM has a higher temperature spec compared to ReRAM, giving MRAM an edge in applications like automotive. “The short answer is MRAM can be used for the majority of auto applications, but ReRAM can only be used in consumer applications today,” UMC’s Uriu said.

Fig. 6: MRAM vs. ReRAM Source: UMC

So far, Adesto and Panasonic are the only companies shipping standalone ReRAMs. Crossbar also is working on standalone devices, although the company is focusing on an IP licensing model. In the embedded space, Crossbar is working with Microsemi. “Microsemi is working to integrate embedded ReRAM into an advanced SoC or FPGA. That is in 14nm or 12nm,” Crossbar’s Dubois said. “That means ReRAM scales. We have a foundry that is working with us to get 12nm ReRAM in their mass production line.”

Others are also developing ReRAM. “Everybody is looking at ReRAM right now,” Dubois said. “If you want read-intensive applications, ReRAM is much better.”

For embedded ReRAM, the main apps include AI/machine learning, computing, home automation, industrial and security.

Other contenders
Ferroelectric RAM (FRAM) is another technology to watch. Using a ferroelectric capacitor to store data, FRAM is a nonvolatile memory with unlimited endurance.

Traditional FRAMs are limited in terms of scaling. To solve these problems, startup Ferroelectric Memory (FMC) is developing a next-generation FRAM, dubbed a ferroelectric FET (FeFET).

Still in R&D, a FeFET isn’t a new device. A FeFET makes use of an existing logic transistor with a high-k/metal-gate stack based on hafnium oxide. The gate insulator is then modified with ferroelectric properties.

“What we are doing is mainly a transistor-based technology, a one transistor-based ferroelectric memory,” said Stefan Müller, chief executive of FMC. “We are pushing forward in the embedded domain. This will be something in the consumer segment.”

Meanwhile, in R&D, Nantero is developing carbon nanotube RAMs. For embedded apps, Fujitsu is expected to offer the first carbon nanotube RAMs based on Nantero’s technology.

“The strategy is to do embedded memory for logic. Fujitsu will be ramping that in 2019,” Nantero’s Doller said. “In the meantime, what we are working on is a DRAM-compatible high-capacity device. That will compete with DRAM.”

So the next-generation memories are making progress, giving OEMs plenty of options. But they still have a long way to go before they are mainstream devices. They may never reach that point, as DRAM and flash continue to roll along.

Related Stories
Four Foundries Back MRAM
What Happened To ReRAM?
A New Memory Contender?
Embedded Flash Scaling Limits
New Embedded Memories Ahead