Cheap, fast, durable, energy efficient, and non-volatile, retaining memory even when the power is switched off (or fails). Those are the requirements for universal memory that would melt the present trio of magnetic hard disk, DRAM, and flash memory.
Today's computers are rather awkward. Most still have spinning hard-disks, which have the advantage that they are cheap and have more capacity than most users need, but they are also slow and vulnerable, and take up a lot of space. There already is a replacement for that in the form of flash memory-based 'solid state drives.'
These have many advantages over hard disks. They are smaller, much less vulnerable (there are no mechanical parts moving), a whole lot faster (enabling instant-on computing), and they consume a lot less energy, which is especially helpful in mobile devices. If you really want to boost your computer, change your hard disk for a solid state drive. The difference really is remarkable.
But, alas, flash memory has disadvantages. It's expensive, it doesn't last forever (hence the need for software to balance the load to prevent some parts from wearing out much faster than others, and other software to keep an eye on degeneration of cells).
While flash is a lot faster than spinning hard disks, it's a lot slower than DRAM (dynamic random access memory), the stuff which is like the 'working memory' of computers. DRAM is fast, cheap, and durable, but alas, it's volatile. The content disappears after the power is turned off (or fails), without scope for recovering it.
It would be nice to have memory that combines the best of these three. Be as fast, energy efficient and durable as DRAM, as cheap and capacious as spinning hard disks, and as non-volatile as flash. Actually, such wonderful memory might be closer than you think. First, why DRAM is so efficient:
DRAM uses a small capacitor as a memory element, wires to carry current to and from it, and a transistor to control it - referred to as a "1T1C" cell. This makes DRAM the highest-density RAM currently available, and thus the least expensive, which is why it is used for the majority of RAM found in a computer. [wikipedia]
But alas, it's volatile.. So here are a few of the candidates for the next generation of computer memory and the important players behind this.
Ferroelectric-RAM or F-RAM is similar to DRAM, but it retains data after power is switched off. It's also faster than flash memory but, as of yet, it can't compete with flash on storage capacities and cost.
The players here are Ramtron International (RMTR), the (fabless) company most responsible for its development. Other companies are Texas Instruments (TXN), Rohm (ROHCY.PK), Toshiba (TOSBF.PK), Infineon (IFNNY.PK), Seiko Epson (SEKEY.PK) and Fujitsu (FJTSY.PK). Fujitsu is the company that produced the Ramtron chips until 2010, after which Texas Instruments took over. There are already products on the market for some time, but these have limited uses (like smart cards and really specialist niches) due to their low capacities.
There is progress in the labs, for instance Toshiba's fast 16Mb chip, but there is still quite a way to go. Not interesting enough yet to place bets, we would argue.
Much more promising is Resistance Memory or ReRAM, based on materials whose electrical resistance changes when a voltage is applied and, crucially, keeps this change even when the power is turned off. The advantages of ReRAM are:
- It can read/write at low voltages and high speed, making it as fast as DRAM, about 10 nanoseconds (much faster than flash)
- It is non-volatile, it retains the content after power is turned off
- It doesn't need vacuum conditions in production, making it cheaper than flash
- It's more durable than flash. The latter has write cycles in the ten or hundred thousands, ReRAM has write cycles in the million
- It can also work as 'memristors' (see below)
The companies involved here are Elpida (ELPDF.PK), Sharp (SHCAY.PK) and Panasonic (PC) from Japan, HP (HPQ), and Hynix Semiconductor (HXSCL.PK). Sony (SNE), together with Micron Technology (MU), is aiming to make the technology commercially available over the next three to four years.
Significant developments are still taking place. For instance:
Researchers at University College London have created a new silicon oxide version of ReRAM that's about a hundred times as quick as good old NAND storage while using one thousandth of the energy - and doesn't require a vacuum to produce, meaning it will cost less to produce. [TrustedReviews]
The breakthrough at UCL was that they found a way to make the chips without the need for a vacuum, which makes the chips more durable and cheaper to produce, and produced chips that are capable at operating at ambient temperatures. This could be quite a breakthrough as
Previous ReRAM prototypes, including a headline-grabbing example from 2008 which was based on titanium dioxide and Elpida's recently-announced module which is due for mass production in 2013, have required esoteric operating environments - including vacuums and extreme temperatures - and complex production processes which have precluded them from commercialisation. [bit-tech]
Magnetoresistive Random Access Memory (MRAM) uses magnetic storage elements to store data instead of storing it as an electric charge like flash memory (see here for a technical description). The advantages:
- Much faster than flash memory
- Much longer life cycle
- Non-volatile, retains data after power is turned off
However, this isn't yet ready for taking over as the capacity of the chips is low. Freescale (FSL) released a 4Mbit chip last year (costing a whopping $20+). However, things might speed up with IBM (IBM) and TDK (TTDY.PK) joining forces, and Freescale is also betting big on the technology (it has licensed it to Honeywell (HON) already).
Products nevertheless are on the way like this combination product from Buffalo from Japan, which combines a (now) traditional flash based solid state drive with MRAM cache memory. Still, MRAM isn't ready yet for a prime time assault on computer memory or replacing flash.
There are variants of MRAM like the orthogonal spin transfer magneto-resistive RAM (OST-MRAM) proposed by start-up Spin Transfer Technologies. The technology was developed by Dr. Andrew Kent from New York University. The story is spun nicely in various places but they did get $36M in new capital to
Spin Transfer Technologies (STT) will use the funds to scale operations, purchase equipment and grow its team to accelerate development of its patented orthogonal spin transfer magneto resistive random access memory technology (OST-MRAM). The company is poised to create the next generation of memory applications combining the non-volatility of flash with the read and write performance of DRAM and SRAM into one, seamless product. Initial performance data from STT bit cells has far exceeded industry standards in key areas.
Promising, but not yet ready to assault flash either.
PCM or PRAM
Phase Change Memory, or PCM, takes advantage of the alterable physical and electrical properties of certain materials in order to store information, similar to CD's or DVD's. It is developed by Intel (INTC), Numonyx, a joint venture by Intel and STMicroelectronics (STM) and was bought by Micron; and Samsung (SSNLF.PK)
At least in the recent iteration of IBM, has a few major advantages over flash:
- It's 100x faster than flash memory (although less so in practice, but still faster)
- It is 'bit alterable', which means that unlike flash, stored information can be switched from one to zero, or zero to one, without a separate erase step (which explains a good deal of the speed advantage over flash)
- It's much more durable, millions of write-cycles are possible before wearing out (at least 10x more durable than flash)
- It's non volatile, retaining data after power is switched off
- It's also scalable and stackable, so future proof (flash memory has problems here due to the floating gate memory structures which are difficult to shrink)
However, problems exist, like temperature sensitivity, and capacities can't yet compete with flash (although they're getting closer). Apart from IBM, Micron already has products. Samsung is perhaps the front runner. While cheap in principle, it can't yet compete on cost (which is normal for a non-established technology):
The biggest challenge facing PCM is cost, said Handy. The first-generation PCM chips are more than twice as expensive as established chips such as DRAM and flash due to poor economies of scale and limited R&D, he said. PCM may be unable to compete on price with NAND flash, which is used to store images and movies on devices such as smartphones, said Handy, who believes NAND could have enough of a price advantage to hinder PCM adoption. [enterprise storage forum]
This technology is getting closer to prime time, so we would keep an eye out for Micron and Samsung in particular, with a product out already. This is an 8GB (20nm) chip, so capacity-wise it can't compete with NAND flash, but the density is already right up there with the best in flash memory. There are other advantages:
First of all it uses a much simpler design with fewer process steps, which potentially makes it cheaper to manufacture at high volumes. It also has a much better endurance as it can tolerate hundreds of thousands of writes compared to flash that has a tendency to wear out after less than 10,000 writes. Add to this that the chip is bit addressable and has a bandwidth of 40MB/s and is compatible with NOR flash. This makes the chip a potential candidate for hybrid memory systems in the initial phase, but could eventually lead to memory system based solely on PRAM. [Semiaccurate.com]
It could be an evolutionary step from flash when that technology reaches the limits of density.
Spin-transfer torque RAM, or STT-MRAM, uses the spin of electrons to change the magnetic orientation of a magnetic layer (effectively creating a zero or one; if you must know, here is a technical explanation). This kind of MRAM memory was developed by Grandis (which was acquired by Samsung in August 2011).
Hitachi (HTHIY.PK) also demonstrated a (32Mbit) prototype in 2009, as did Qualcomm (QCOM). There are other (non-public) companies working on this technology, like Everspin Technologies, Crocus Technology and Spin Transfer Technology.
According to pioneer Grandis, the advantages are:
- Non volatile
- Highly scalable
- Low power consumption
- SRAM read/write speed
- Unlimited endurance
- DRAM and Flash density
- Multi-level cell capability
Products from Grandis (even before the Samsung takeover) and Everspin were announced, albeit in niches like embedded SRAM and low power mobile RAM replacement. The initial market for these products, according to Grandis
STT-RAM has key initial markets replacing embedded technologies such as eSRAM, eFlash and DRAM, and providing new functionality at 65 nm and beyond. In automotive applications, it has higher speed and lower power than eFlash and is denser than eSRAM. In portable and handset applications, it can eliminate multi-chip packages (MCPs), provide a unified memory subsystem, and reduce system power consumption for extended battery life. In personal computers, it can replace SRAM for high-speed cache, Flash for non-volatile cache, and PSRAM and DRAM for high-speed program execution.
There are still hurdles to overcome though:
Further increasing the storage density remains a challenge, however, because the write current needs to be increased to keep the bit thermally stable. [nanowerk]
Keep an eye on Hitatchi and Samsung in particular. Toshiba is already planning to use this type of memory as cache memory, adding it between DRAM and NAND flash. But this isn't replacing flash and it's a niche market.
Complementary Metal Oxide memory is a resistance-change memory (see here for a technical explanation). The advantages, according to the main developer of the technology, startup Unity Semiconductor Corporation, are:
- 4x the density of NAND flash
- 5x-10x the write speed of NAND flash
- Scalable below 20nm
The important player is Unity, which is a fabless developer, in cooperation with Micron Technology. Rambus (RMBS) has acquired Unity last February for $35M in cash. They (obviously) see a strong future ahead for this technology:
According to Rambus, speaking with TechEye, NAND has its days numbered due to the inherent difficulties in scaling the cells below a certain node (20nm), precisely where CMOx leverages its advantages. Calling it an "inflection point" in NAND usage, Rambus is betting on CMOx to carry Flash smaller nodes. At 17nm, CMOx will have four times the density of NAND. [Techeye]
Products were targeted near time when Unity was independent. From 2009:
The company has targeted a 64 Gigabit class of NVM (non-volatile memory, which is 8GB per chip). It's described as a "passive rewritable cross-point memory array" with no transistors in the memory cell, and the company has 64 Megabit products (8MB per chip) in production for one year. The 64Gb chips are "close to tape-out and slated for pilot production in 2H'2010, with volume production in 2Q'2011″. [Geek.com]
However, Rambus isn't racing to market, though:
A timeframe for the mass-market of CMOx has not yet been set by Rambus, but considering the technical requirements and the current state of manufacturing, within two years Rambus will have something that is able to compete with NAND, we gather. Right now the goal is to "broadly licence the technology and create an ecosystem of tools and equipment suppliers to enable the commercialisation of CMOx". [Techeye]
This push-back of actual commercial product timelines is by no means a rare event. In fact, it's more rule than exception, and understandable as flash memory itself (the memory that these alternative technologies try to replace) has seen strong improvements and adoption. The latter have reduced costs considerably, making flash a moving target for the wannabe replacement technologies. We'll keep an eye on Rambus nevertheless.
Conductive Bridging RAM, or CBRAM, is a variant of ReRAM (for the technical workings see here). Developed by Arizona State University and Axon Technologies, the latter has licensed the technology, amongst others to start-up (private) developer Adesto. Adesto, together with foundry partner Altis Semiconductor, announced plans to sample last year:
Ed McKernan, director of business development for Adesto, said: "Yes we will be shipping our first product a 1-Mbit serial EEPROM drop-in replacement in 130-nm CMOS in 2H 2011. With regards to samples; as we had planned, we received internal samples in Q1 and we are sampling customers in Q3." Altis currently manufactures a 130-nm CMOS process at its 200-mm wafer fab with plans to go 90-nm. [EEtimes]
The figures involved (1-Mbit, 130nm) suggests these products aren't terribly advanced and will be niche products. The technology does have advantages, though:
- It is non-volatile
- It consumes 100 times lower current than current flash memory technologies in the market [Adesto in EEtimes, see previous link]
- It is low cost
- It is scalable below 20nm
But Adesto isn't alone (and has non exclusive patents licensed from Axon Technology), big players like Infineon, Micron, Sony, and NEC are also working on their own variants of the technology. NEC has a variant called 'Nanobridge'.
It doesn't seem that the technology is ready anytime soon to replace flash on a substantial scale, though.
This is a type of memory that stores data as magnetic patterns on tiny wires (for a technical description, see here). IMB is the big player here, but the memory seems still quite a bit away from commercial application. It promises more durable, faster memory and IMB did have a sub 10nm prototype.
A potentially even more revolutionary product would be to combine the transistors of processor chips with memory in a single device, allowing chips to both process and store data on the same chip. Today, these tasks are done by separate entities (DRAM and hard disks/flash for the volatile and non-volatile storage parts, computer chips with transistors for the processing); the constant back and flow of data slows things down and wastes energy.
Memristors, first proposed in 1971 by professor Chua, are the fourth basic building block of electronic circuits, after capacitors, resistors, and inductors. This technology would enable computers that would be more akin to the human brain:
Researchers at the University of Michigan recently showed that the devices can mimic synaptic activity in the brain. [BBC]
They are also stackable, that is, layers can be build upon one another creating chips with a really ridiculous capacity. Apart from that, they can also be shrunk way beyond what transistor based chips can be:
Professor James Tour of Rice University in Houston said the memristor's ability to be compatible with existing transistor based technologies was a "critical parameter to permit rapid implementation into present chip manufacturing processes". Dr Williams said he had already made "crude" prototypes with features as small as 3nm. "The functional equivalent of Moore's Law could go on for decades after we hit the wall where we can no longer shrink transistors," he said. [BBC]
3nm (nanometers) -- that is way smaller than the 22nm circuits that are presently in the forefront of commercial chips (which Intel alone has achieved). HP is at the forefront of memristors and it has teamed up with Hynix (HXSCL.PK) in September 2010.
This article wouldn't be complete (which it isn't anyway, needless to say) without mentioning the target that these replacement technologies try to replace. While the inherent technological properties might place flash at a disadvantage versus most of the alternatives above, flash has the significant advantage of being a commercially proven technology with a rapidly increasing market.
This drives both manufacturing cost advantages (economies of scale, learning) and product advances, like Samsung's whopping 10nm NAND flash memory. In technology markets, one often sees a last-dash spurt in technological advancement of an established technology when it comes under siege from supposedly better alternatives; flash doesn't seem an exception to this phenomenon.
Here is what a panel at Applied Materials (AMAT) concluded:
The panel consensus was pragmatic: evolutionary advances in conventional technology, coupled with fast-maturing methods of stacking multiple die in a single package, can meet performance, power and cost targets for the next several years. [Applied Materials]
Some people on the same panel also concluded that the quest for universal memory is a 40 year old myth and that the physics are against it. While they might be right, we think it's useful to keep an eye on the developments in this space, as the potential pay-off is huge. We're especially interested in the recent advances in ReRAM at University College London.
None of the alternatives mentioned here seem ready for a prime-time assault on flash memory in the near future, as the latter keeps improving and has the advantage of having an established production base, reaping stable yields, and economies of scale and learning. However, most of the alternatives have inherent advantages. The route to market seems to be to gather production experience via niche markets to see whether that can drive advancement fast enough to expand the applications.
In any case, it's good to know that when the advancement of flash slows down, and/or when some company puts its full weight behind one of the alternatives, the market can take off in earnest. Samsung seems pretty well placed to do that, but they are also the biggest player in flash memory so they might want to ride that wave first.
We are especially curious to what seem to be a rather dramatic improvement in the ReRAM memory from University College London. Were will that technology end up??
Disclosure: I have no positions in any stocks mentioned, and no plans to initiate any positions within the next 72 hours.