RRAM: Now Used in IoT, AI and Data Center Storage

Download this presentation: RRAM: Now Used in IoT, AI and Data Center Storage

00:02 Tom Coughlin: Hello, I'm Tom Coughlin. I'm the president of Coughlin Associates, and I'm glad to be speaking here at the 2020 Flash Memory Summit.

Coughlin Associates is my company. We've been doing consulting for over 20 years now in digital storage and memory and their applications, various types of work. We also write things. I have a blog I do on storage and memory at Forbes.com and other websites as well, published reports on digital storage and memory technologies and their applications, and in fact, one of those will be the basis of much of the material I will present today. And I also have been putting on storage- and memory-oriented events, including the Storage Visions Conference, the Creative Storage Conference, and other events as well. And, in fact, I was general chair of the Flash Memory Summit for 10 years.

00:52 TC: Today, I'm going to talk to you about resistive random access memory, and the situation today and what's going on with this memory technology. Here's an outline of my talk. We'll talk first on the properties of various emerging memories, including resistive RAM. We'll talk about some company announcements regarding resistive RAM memory. We'll talk about the technologies themselves, because resistive RAM is a term that actually incorporates a number of different memory technologies. Then we'll talk about some advantages of resistive RAM, as well as the operation and construction of resistive RAM memory devices, some applications for resistive RAM, and then because TSMC, a major foundry, has been promising to come out with resistive RAM products this year, we'll talk a bit about the recent updates that they have looked at or worked on. And, finally, we'll end up with a summary.

01:49 TC: Let's do a comparison of various solid-state memory technologies. If you look at this table here, it's from the report that Jim Handy and I have done on emerging memory technologies, "Emerging Memories Find Their Direction." And we have here on the left-hand side, some established memory types, including static random access memory, dynamic random access memory, NOR flash and NAND. These are in higher-volume production. And then more emerging memory types on the right-hand side here, which are in lower levels of production or, in some cases, perhaps still in pre-production or very low-volume products. Now, these are both standalone as well as embedded memory products potentially, as are some of these established memory types.

02:35 TC: And the emerging memories we're talking about are ferroelectric random access memory (FRAM); resistive random access memory, which is initialled as ReRAM or RRAM; a type of magnetic random access memory called toggle mode; another type of magnetic random access memory called spin-tunnel torque MRAM; and then phase-change memory. And this last one actually is . . . There's a product called 3D XPoint memory -- which Intel and Micron developed -- which currently is being sold by Intel as their Optane memory technology both in SSDs and in DIMM memory modules.

You can see some various properties about these established memory types and the non-volatile memory types. The non-volatility, that is the data stay on when the power is off, the size of the cells, the read time, the write time, endurance, the number of times the memory can be rewritten, how much energy is being used to write, and also what kind of voltage is required for writing. And you can see some comparisons here between these established memory types and emerging memory types.

03:38 TC: So, there's issues on these established memory types. SRAM is fairly large, takes up a lot of real estate for a memory cell. DRAM is totally non-volatile, has to be refreshed fairly often, uses a fair amount of energy. NOR flash is perhaps reaching limit . . . seems to be reaching limits and how much you can scale it to less than 28 nanometers. And NAND flash, of course, which is the major non-volatile memory technology used in SSDs, and we'll compare those . . . These can be compared then to the emerging memory types. You could see there's various advantages there. And resistive memory, promises fairly low-sized cells -- that means you get high density, pretty good read speeds, write time is pretty quick, and it also can have reasonable endurances for some applications, and the write energies are low. So, there's some good things about resistive RAM, and we'll talk a bit more about that.

04:41 TC: Let's look at some of the companies, in fact, that have been promoting resistive RAM memories. There's a company called Crossbar, which in 2016 announced a partnership with Semiconductor Manufacturing International Corporation (SMIC), which is another major foundry, to codevelop and produce resistive RAM technologies for use in MCUs, system on chips for IoT, consumer electronics, artificial intelligence, and industrial applications. They developed their own . . . Crossbar has developed its own resistive memory technology which they believe they can build Crossbar structures with and therefore get multiple layers of memory cells. However, to date, no actual products from Crossbar or SMIC with resistive RAM have been reported.

05:28 TC: Another case is Panasonic. In 2017, Panasonic announced a partnership with United Microelectronics Corporation (UMC), which is another one of these semiconductor foundries, to codevelop and produce resistive RAM devices initially with a 40 nanometer process, which in 2017 wasn't too bad. Products were supposed to be ramping in 2019 for microcontrollers, but again, no products have been reported to date.

06:00 TC: Another case is TSMC, which I hinted at before, and I'll be talking about a bit more later. At a technology events in August of this year, TSMC said that it is offering resistive RAM, embedded non-volatile memory on its 40 nanometer and 22 nanometer manufacturing processes. The company said that multiple customers would tape out ICs with embedded resistive RAM in the second half of 2020. They indicated that Infineon Technologies AG has adopted resistive RAM for certain specialized microcontrollers for the memory in microcontrollers, so it seems like the biggest progress possibility for a product soon seems to be coming from TSMC.

06:47 TC: Now, another company, which actually has been making standalone products, is Adesto, which was acquired by Dialog Semiconductor. That acquisition finished in June of 2020. They have been shipping their CBRAM, which is a resistive RAM technology, as standalone memory chips for several years. In October of 2020, Dialog Semi said that they would license -- this is supposed to be the CBRAM memory technology -- to GlobalFoundries, another foundry company, semiconductor foundry company, which also has a deal with Everspin licensing their MRAM technology. And GlobalFoundries will first offer Dialog's CBRAM as an embedded non-volatile memory option on its 22FDX platform with the plan to extend to other platforms.

07:40 TC: Now, let's talk about some of the different resistive RAM technologies that are out there, and there are several. And in a sense, you'd say that many of the other memories, which have different names are . . . Their memory is by changing the resistance of the memory cell. For instance, MRAM basically changes the memory cell resistance as a way of storing information. But this is a particular set of . . . The ones we're going to be calling resistive RAM here are particular set of technologies, one of which is the filament devices.

08:14 TC: There's two basic types of resistive RAM technologies. Below, we are going to talk about another technology as well. The first is metal filament resistive RAMs. And they function through the formation or dissolution of a metallic filament or nanobridge in a conductive electrolytes using electrochemistry. In the second type, the binary oxide device, conduction is caused by a reversible filament formation involving oxidation and reduction. A subset of this mechanism, referred to as interfacial switching, is caused by oxygen vacancy drift diffusion that induces barrier modulation. The metal filament resistive RAMs, as shown here, switches through the formation of a conductive filament. The filament composition is metallic or has near-metallic conductivity. It's often referred to as a conducting-bridge device. This figure in the slide here shows the method for this formation growth and modulation by an applied voltage.

09:16 TC: Current control during the filament forming and during the set process determines the filament size as well as the diameter and, therefore, the operational characteristics of this resistive RAM device. Write currents on the order of 20 to 100 microamperes have been reported. Test results of resistive RAM arrays indicate switching times in the order of 10 nanoseconds and endurance of 10⁵ to 10⁶ cycles. A significant advantage over flash, whose program and erase times takes tens of milliseconds, which can only endure 10² or 10⁵ cycles. So, this is a better endurance than NAND flash.

09:55 TC: Another type of resistive RAM device consists of two metal electrodes between which a unique solid dielectric or electrolyte is dispersed. This dielectric material is usually a transition metal oxide. Shown here, it's a tantalum oxide. Dielectric materials that have been used are titanium oxide, nickel oxide, hafnium oxide, tungsten oxide, tantalum oxide, vanadium oxide, and copper oxide, as well as several others. The top and bottom electrodes are made out of platinum, titanium nitride, ruthenium, or nickel. And a wide variety of combinations of these electrodes and dielectrics is reported in the literature. However, the use of a titanium nitride bottom electrode in hafnium oxide dielectric is considered an attractive configuration of a simple material system that yields good switching characteristics. This type is typically referred to as an oxygen vacancy-type resistive RAM device.

10:58 TC: In this figure, two layers are positioned between two electrodes. The lower one is of tantalum oxide and the upper one of Ta2O5minus delta. The binary oxide resistive RAM device largely consists of oxygen vacancy technologies. The fundamental switching mechanism for oxygen vacancy resistive RAM also involves the formation of a filament. But this filament consists of oxygen vacancies within the dielectric between the metal electrodes, the results from the application of a higher set or Reset voltage.

A very small gap exists between the filament and top electrode, and filament conduction to the top electrode has been postulated to be due to tunneling through the dielectric. Performance of this type of resistive RAM is thought to depend on a type of electron flow rather than ionic current conduction after filament formation. And this allows the technology to attain switching speeds equivalent to or better than flash memory in much lower right currents.

12:04 TC: Another approach to resistive memory is called CeRAM for correlated electron random access memory. It was pioneered by a company called Symetrix, which is a ferroelectric intellectual property development company.

CeRAM uses similar materials to a resistive RAM, but employs a much thinner transition metal oxide layer, as you could see here, to convert the material's operation from a filamentary to a phased transition memory. Rather than using oxygen diffusion to rupture and reform filaments, the technology uses what's called a Mott charge transfer phase transition to create conductive and insulating states. The company says that this creates a more stable and repeatable memory element with all the other benefits of certain resistive RAMs, including the use of silicon-friendly materials.

In October of 2020, Arm, which had acquired the Symetrix CeRAM IP, transferred that IP portfolio and related patents to a new spin-out from the company Cerfe Labs, headquartered in Austin, Texas, and took a minority ownership stake in the new company.

13:14 TC: Now, Cerfe Labs was recently announced as taking the CeRAM technology. And as part of the spin-out, Arm transferred its IP portfolio, more than 150 patents, to Cerfe Labs, which should be the foundation of a roadmap for related CeRAM technologies. The initial focus will be on producing meaningful prototypes, which will be licensed to partners with the goal of accelerating timing to enable those novel non-volatile materials for systems applications.

13:49 TC: Now, we talked a bit about some of the companies that are out there. Now, let's talk about the operation of resistive RAM itself. And one of the things about resistive RAM that's really interesting is the scaling. Reducing the scale of resistive RAM cell is believed to actually increase the on-and-off resistance ratio due to the fact that the on current, the current flowing through the metallic filaments, does not shrink with scaling. While the off current, that is the cell's leakage current, which is the function of the cell area, decreases as the cell size shrinks. That means that resistive RAM device characteristics should improve as the device is scaled. In other words, the resistance difference between the on-and-off state gets larger, as you could see in the chart shown here.

14:35 TC: This scaling effect is expected to enable even further cost reductions since the increased read margins allow multiple bit levels to be stored in each cell the same way as is currently done in MLC and TLC, and nowadays even QLC flash memory. This figure shows the scaling of resistance with cell size. Resistive RAM is estimated to be scalable below 10 nanometer process that rules. Today's write currents for a metal filament resistive RAM is at least 10x higher than flash, 20 milliamperes versus 2 microamperes. Although write time is less than 10 nanoseconds compared with about 15 milliseconds for flash, or 5 million times shorter, filamentary resistive RAM poses unique manufacturing challenges. The cell resistance has an inherent variability that is proving difficult to manage using today's wafer processing techniques. Additionally, the chalcogenide glasses used in some of these technologies are difficult to integrate into a standard CMOS process.

15:44 TC: Here, we're showing a little bit about the operation of these devices here. You can see the basic device, the top electrode, the bottom electrode, and the middle oxide between. This is the oxide type. The current levels and voltages required for resistive RAM switching are generally shown in the figure here on the right-hand side.

For either metal bridge or oxygen vacancy resistive RAMs, the bit cell consists of a resistive RAM structure deposited above the select device at the back end of the line, and the cell dimensions will be determined by the resistive RAM or select device, whichever is larger. Now, the select device keeps the current from this device from leaking over into other memory cells, so it serves a very important function.

16:27 TC: Here's looking a bit at how resistive RAM is integrated into CMOS. In one technique for resistive RAM-CMOS integration, resistive RAM layers are deposited on top of the preprocessed CMOS logic wafers, which you can see shown here. The resistive RAM cell is added between two of the metal layers that are deposited in the later back end, that's the top part of the wafer processing. This approach is illustrated in this image. The process can be achieved with a very small number of additional mask steps. Adesto, for instance, claims that its embedded resistive RAM process can be built using no more than a single mask at upper metal layers and with only two masks at the first lowest metal layer.

17:16 TC: Here's a two-mask resistive RAM element in tungsten vias, for example. This image illustrates conventional tungsten contact at the right. And a simple CBRAM conductive bridge random access memory bit cell inserted within that contact on the left.

So, you see a regular V on the right, and then the MRAM inserted on the left. According to Adesto, this memory element can be produced using only one or two mask layers, reducing the overall cost of the wafer when compared to the higher mask count required, for instance, by NOR flash. The embedded photograph that you see in the middle is the same structure when seen through a tunneling electron microscope. Adesto says this technology then could be easily integrated into a CMOS logic process because it resembles the construction of a standard metal insulator, metal capacitor that is widely used to make tiny capacitors in back-end foundry processes. The thermal processes thermal budget is also said to be more amenable to CMOS logic than are either MRAM or FRAM, that's the ferroelectric RAM.

18:32 TC: Now, let's look at some of the ways in which people are talking about using 3D RAM. This is a 3D structure which is using techniques similar to what we're doing at 3D NAND to other memory types to drive up chip densities and substantially reduce costs. A diode selector resistive RAM, like flash memory, can be built as a 3D . . . to help build a 3D memory device, as shown here. In this design, multiple layers of memory share the same set of lithography steps to significantly reduce processing cost. If a 30-layer structure could be made, an 8F2 footprint would be reduced to an effective cell size of 0.25 F2. This 3D design uses the internal selection mechanism of the crossbar memory to allow multiple devices to share a transistor selector.

19:25 TC: Now, here's a stack cross-point array design here. Since resistive RAMs are set and cleared by using currents in opposite directions, many are controlled by bidirectional diode selector. This figure shows resistive RAM stack cross-point array with such a selector.

First, in the unselected state, each cell must be stable during the total operation. As a programming voltage is applied to a specific word line, a half-select voltage is applied to the adjacent word lines. This half-select voltage helps maintain stability for millions of successful read/write cycles. In the selected state for read and write, the desired current or pulse passes through a specific device only. The most promising resistive RAM structures involve multifilm diodes using compatible materials or other two-terminal selector types as an integral part of the unit cell in order to eliminate the requirement for a larger selector such as a transistor, which would then make the memory cells much larger. Considerable development work is being directed to implementing the space-efficient structure, and without this advanced resistive RAM density potential is probably no better than existing technologies that use the transistor as a select device.

20:41 TC: Now, here's some interesting examples here. While most resistive RAMs use a diode or bidirectional diode selector, one company's product stands apart from others because it does not need a selector. This is Crossbar Technologies that developed a metal filament cell that performs like both a memory bit and as a diode. During programming, the Crossbar cell does not form a complete metal filament, but the filament comes within a few atoms of bridging the gap. When a forward read current is applied, the bridge is completed and the cell exhibits a low resistance. But when the current is removed, the last two atoms again disappear. If a reverse current is applied, the gap causes the cell to be in a high resistance state. So, it acts as its own diode.

21:26 TC: The Crossbar cell requires a specific voltage range for programming that is greater than, plus or minus, one volt. Programming is only possible for voltage levels beyond this range. Care must be taken not to exceed the breakdown voltage of the oxide insulator or medium. This topology is not yet in production. The company is ramping it's resistive RAM into volume by first using a more conventional 1T1R cell structure with a self-selecting memory to ramp after the basic technology has reached volume. The 1T1R cell is aimed at embedded memories and SoCs, while the self-selecting cell will be used in a 1TNR configuration for the large memory arrays found in standalone memory chips.

22:11 TC: In the quest to improve the bit capacity per chip, as well as to minimize space requirements, a 3D RAM structuring technology is being actively pursued by many developers. It is estimated that a 3D RAM non-volatile technology would have an estimated greater than 1 terabyte on a single chip with existing lithographies. Better specifications in flash, about 100x lower latency, 20 times faster writes, small pages, and no block-erase requirements, 20 times lower power per bit, and manufacturing using fully compatible as an addition to CMOS processing, and scalable to higher densities and capacity. And with the use of a 3D resistive RAM, a high-capacity, dense, low-latency, non-volatile memory is possible that is fully compatible with the CMOS standard manufacturing process. And that's what would be shown in this figure.

23:05 TC: Now, another interesting thing about these resistive memory technologies and, in fact, in the larger scale, resistive RAM as well as phase change, what's called phase-change memory, is that they're being extensively investigated for use in artificial intelligence applications. In particular, they lend themselves to use in linear waiting configurations in an architecture, commonly called a neural network, and used in neuromorphic computing. Neural networks are very simplified type of inference engine that can perform an enormous amount of math at low precision in a very short time. A simplistic view of a neural network appears here, and here's an article on the right-hand side that happened recently that was covering using, in this case, resistive RAM material, they call it memristor, for neuromorphic computing.

24:03 TC: Now, we talked a bit about company announcements, we talked about the operation of various types of resistive RAM. We indicated how these are being integrated and potentially used. TSMC, though, deserves some particular attention here. In a recent technology seminar that they did in late August, they talked about the use of resistive RAM. In fact, you could see it as part of their technology or their integrated specialty technology platform for system-level solutions under NVM, non-volatile memory, and includes traditional NOR flash, but it also includes MRAM and resistive RAM. And, in particular, let's look at the way they're using this there.

24:52 TC: They said that their 40 and 22 nanometer resistive RAM is ready for production. It's eFlash alternative for IoT in smart card and those that can replace NOR flash and extends to support 10-year retention at 125 degrees after 10,000 cycles. So, this has to do with that higher endurance. And they said that there'd be multiple customers which would do tape-outs in the second half of 2020, which means that initial products could be available sometime in early 2021.

25:32 TC: They also talk about MRAM, which is not, of course, the major topic of this talk. A 22 nanometer MRAM is ready for production, automotive qualification of Q4 2020, 16 MRAM targets risk production for eFlash-like in the fourth quarter of 2021, and RAM-like in the fourth quarter of 2022, that's DRAM-like. And, in fact, there are currently some products that have been announced that are manufactured in TSMC's foundry, for instance, AI applications of MRAM.

As we indicated before, apparently Infineon is looking at microprocessor applications using TSMC's resistive RAM. Part of their announcement, they talked about bipolar-CMOS-DMOS -- BCD they call it in their power management process -- and power management, use in power management integrated circuits where the resistive RAM is used as their non-volatile memory.

26:34 TC: And you can see that they show this on a roadmap here with the 40 nanometer process, 12 volts, 28 volts with resistive RAM. They got various applications that they show here, including automotive and consumer applications. Also, here's another picture of this power management integrated circuit technology platform. You can see the resistive RAM shown here with the other parts of the technology. Target application, higher digital content, and again they incorporate the resistive RAM for their embedded non-volatile memory.

27:16 TC: In summary, resistive RAM technologies have long been available in discrete memories, companies such as Adesto. A resistive RAM has advantages in scaling versus other memories, such as NAND flash, and there's been even talk that when NAND flash does run out of gas in terms of 3D stacking and lithographic dimensions, that resistive RAM may be its replacement memory because it can scale lower in terms of lithographic dimension. And although it appears to be lagging MRAM at embedded applications, recent announcements indicate that resistive RAM will be used soon in several embedded applications for consumer, industrial, and other uses.

27:56 TC: And I'd just like to draw your attention, a lot of the material from this presentation came from a report that Jim Handy of Objective Analysis and Coughlin Associates did, which is "Emerging Memories Find Their Direction," describes the entire emerging memory ecosystem, the technology's phase-change memory, resistive RAM, magnetic random access memory and ferroelectric memories, among others. Talks about the companies, the markets, support requirements, and our forecast, examining emerging memory consumption embedded as well as discrete, and the capital equipment, for instance, required to manufacture MRAM to meet the anticipated demand. It's a 201-page report, with 31 tables and 142 figures. And if you're interested to learn more, you can check out the URLs at the bottom of this slide.

With that, I'd like to thank you for my talk, and I hope that it's been useful to you.