Crossbar RRAM promises to beat NAND speed, endurance

Move over, NAND flash memory. Santa Clara-based startup Crossbar, Inc. has developed a non-volatile memory technology based on resistive random-access memory. Dubbed Crossbar RRAM, the tech promises 20X the write performance of NAND flash memory with 10X the endurance. RRAM isn’t just a distant speck on the horizon, either. According to the Crossbar press release, a working test chip has already been produced. Company CEO George Minassian says Crossbar RRAM is “easy to manufacture and ready for commercialization,” as well. 

Source: Crossbar

The Crossbar tech is based on RRAM patents licensed exclusively from the University of Michigan. Each cell sandwiches an insulating switching medium between electrode layers. Applying a voltage to those electrodes causes nanoparticles in the switching medium to form a conductive filament between them. The simple cell structure is amenable to stacking, according to Crossbar, and it can be scaled down to fabrication nodes smaller than five nanometers.

NAND endurance tends to diminish at smaller cell geometries, but longevity shouldn’t be a problem for Crossbar RRAM. There’s no need to erase the contents of cells before reprogramming them, which should cut down on write/erase cycles. The cells themselves are quite robust, too. Crossbar claims they can withstand one million write cycles. That figure likely refers to single-bit applications, but RRAM also supports MLC configurations with multiple bits per cell. The memory type’s claimed 20-year data retention looks pretty solid, too.

Crossbar touts RRAM’s easy integration with SoC circuitry, and the firm seems intent on providing on-chip storage for smartphone and tablet processors. Its first product will be an embedded SoC. SSDs—and particularly enterprise-oriented drives—are also mentioned as potential applications. 

Source: Crossbar

RRAM’s prodigious storage potential makes it particularly appealing for PC applications. Crossbar says the technology can squeeze up to a terabyte onto a single, 200-mm² chip, though it doesn’t detail the number of layers or the fabrication process required to achieve that feat. For the 25-nm chip pictured above, which is presumably a single-layer implementation, Crossbar RRAM more than doubles the storage density of NAND flash.

Flash memory is big business these days, but its limited write endurance presents a very real challenge to scaling down to smaller process geometry. RRAM looks like an attractive alternative, especially if SoC integration is as easy as advertised. We’ll have to see how long it takes Crossbar to get RRAM products on the market—and how much they cost per gigabyte, of course. If you’re curious about Crossbar RRAM, you can read more about it in this whitepaper (PDF).

Comments closed
    • mikato
    • 6 years ago

    [quote<]Each cell sandwiches an insulating switching medium[/quote<] cell sandwiches, mmm mm mmm

    • HisDivineOrder
    • 6 years ago

    This reminds me of all those times companies threaten to replace rubber tires with something innovative and awesome.

    Add twenty years. More rubber tires.

    I’m also thinking of LCD vs OLED and the promise of the truly low-cost, high capacity SSD.

    Pipe dreams. They’re milking us and they ain’t gonna imperil an entire industry to change course mid-stream. They’re going to buy these guys out, take a few ideas out of the whole thing and implement them into standard NAND, and then quietly shut them down.

    Remember ten years ago when were all dreaming of OLED displays as thin as air and with blacks as inky black as the real thing? But what do we buy when we buy our HDTV’s? What are even the new (so-called) 4K displays using?

    LCD or plasma. The same, tired tech from over 20 years ago.

    More rubber tires.

      • confusedpenguin
      • 6 years ago

      Nuh uh. No Way. You gotta be kidding. That would never happen in real life. These guys will put all the NAND companies out of business and end up being the only player in the game with a superior product, and will sell their product for an affordable price, ya know, for the good of mankind. There is no way that any big evil businesses would intend on buying them out. This whole “we won’t have affordable flying cars tomorrow” attitude is just cuhrazay talk. 😉

      • mikato
      • 6 years ago

      Well as long as there is competition, if the tech is superior in every way then it will survive. Why throw away a way to gain an advantage?

    • davidbowser
    • 6 years ago

    I am shocked and dismayed by the lack of

    “**Yawn** Call me when it’s less than $0.04/GB.”

    comments. Where have the SSD trolls gone?

      • Dashak
      • 6 years ago

      We bought SSDs last year.

        • davidbowser
        • 6 years ago

        Touche

        +1 for that

      • HisDivineOrder
      • 6 years ago

      We’re still waiting for the browser to open.

    • mcnabney
    • 6 years ago

    RAMBUS to buy the patents in 3….2….1….

    • Celess
    • 6 years ago

    Handy Stats from the PDF as Im watching the comment carnage unfold. Should help those not clear on exactly what NAND does as well.

    [code<]Characteristics NAND RRAM Comments Cell Size 5.44F2 4.28F2 provides better array size efficiency Tech Scale <25nm <5nm scales as is nanoparticle-based Program Perf 7MB/s 140MB/s performance is 20x faster Pg/Byte Erase Cant No Need not require Erase prior to Program Byte Program Cant Can greatly improves system performance Async/XIP Read .04MB/s 17MB/s provides faster NAND-like products Prog Engy/Cell 1360pJ 64pJ improves power consumption Cell Endure 3K cyc 1M cyc improves memory system lifetime Retention 1-3 y 20 y provides lifetime and reliability[/code<]

      • Airmantharp
      • 6 years ago

      Thanks for listing that out- but could you provide an explanation for the Characteristics? A couple are a little ambiguous. Thanks!

        • Celess
        • 6 years ago

        Those are basically the same as they are listed in the PDF, with a few left out. Text was shortened but faithfully, so that the columns would flow correctly in the post.

        However:

        Tech Scale:
        here means basically “Process Node” which is a loose definition anyway. In this case they are saying they can go south of 5 nanometer, and NAND can go south of 25 nm, or… we are pretty sure we can get to 5 while we believe NAND wont get much below 25. As an example in the other article it claims the NAND based “840 Series and 840 Pro are made using 21-nm tech”.

        Pg/Byte Erase:
        was ‘page / byte erase’ in the pdf. The ability erase a ‘line’ or just a ‘byte’ at a time. This is another way of saying ‘could it erase enough to do at least a byte’, and they mean lterally ‘at a time’ in the actual array, not via an off (or on) chip controller. In this case they mean the NAND is constrianed to ‘not be able to erase at least a byte at once’ and that RRAM is ‘doesnt need to’. This sounds pretty stupid to say on its ear.

        Byte program:
        ‘write’ version of the ‘erase’ argument above. it says we do writes like we would have done erases, if we actually had to do erases. but, nand …they cant do writes of at least byte or more.

        ——-
        I’m sure if its array/line oriented you could feed it senarios where it would stumble as well. They seem to mostly be talking about performance at a level of what ill call an ‘array’ or one patch of actual memory on a chip. For a long time now, memory like DRAM and nearly everything else, have what amounts to little controllers on the die, like little memory io machines that implement the array management, and marshal the reads, writes or whatever else needs to be done. There can be sets of these on one chip, but generally people have been saying “at the chip level”.

    • OneArmedScissor
    • 6 years ago

    The press release also claims it’s 20x lower power than flash. That could be big for phones. While it can’t replace RAM entirely, it’s yet another building block for new memory arrangements based on non-volatile memory.

    The CPU’s memory channels could be wired directly to non-volatile storage (RRAM, phase change) for reads of the OS / programs / maybe graphics. A few hundred MB of a non-degrading, non-volatile memory (memristors, MRAM) could be integrated into the CPU as a last level cache to handle everything else.

    Flash and eDRAM could technically do that, but they’d chew up phone batteries. There’s a company wiring SSDs into the DDR3 controller for servers, though:

    [url<]http://www.techspot.com/news/53431-diablo-technologies-mcs-architecture-uses-flash-memory-as-ram.html[/url<]

    • 0g1
    • 6 years ago

    Who else read this as “crossbar RDRAM” and thought “oh no, RAMBUS is back”? :p

      • Airmantharp
      • 6 years ago

      They never left…

    • NeelyCam
    • 6 years ago

    Is it me, or does it look like they simply shrunk the NAND die picture to show the nonexisting die pic for their RRAM chip…?

    Pretty awesome that they were able to reduce the size of the I/O circuits at the exact same ratio as the memory itself..

      • ColeLT1
      • 6 years ago

      Cannot unsee.

      • DPete27
      • 6 years ago

      Definitely

      • Grigory
      • 6 years ago

      I noticed that too. Of course it is only to demonstrate the difference in area.

        • NeelyCam
        • 6 years ago

        EETimes has a write-up on this as well:

        [url<]http://www.eetimes.com/document.asp?doc_id=1319139&[/url<] I think the article said that the demo chip they demonstrated was done on a 110nm technology in a 1K by 1K array, but I hit the 2-per-day article limit so I can't double-check.. The listed advantages and the marketing whitepaper make it sound like it's going to be awesome, but I'm a bit unsure about how well they are going to be scaling to 25nm (or 5nm as they are promising). Extrapolating from 110nm is somewhat... shall we say, "prone to inaccuracies"

          • NeelyCam
          • 6 years ago

          The comments on the EETimes article lead me to the website of Symetrix; they claim their FeRAM is 50x faster than flash, and consumes 1/8th of the power at write. The EETimes commenter (C.A. Paz de Araujo) also talks about “CeRAM” that Symetrix has some patent applications on in the system right now.

          Looks pretty interesting, but large-scale commercialization seems just as uncertain as with Crossbar RRAM

    • Bensam123
    • 6 years ago

    That’s pretty awesome… One thing I didn’t see in the news or quickly glancing over the pdf, how fast is it? I think that’s most definitely an important aspect. They go out of their way to point out pretty much every other aspect of the memory, it seems odd that they’d leave that bit out. Is it just a straight drop in replacement for flash?

      • nanoflower
      • 6 years ago

      17MB per sec asynchronous read rate vs 0.04MB per sec for NAND according to their white paper [url<]http://www.crossbar-inc.com/assets/img/media/Crossbar-RRAM-Technology-Whitepaper-080413.pdf[/url<] So it isn't exactly speedy considering they can program it at 140MB/sec vs 7MB/sec for NAND. (edited to exchange the numbers. Had the speeds for reads mixed up as the crossbar technology is much slower than current NAND tech.) LOL. This post has cause me more problems than any other post I've made. It started out with trying to get the URL correct and then got worse. I'll blame it on being horribly sleepy at the time. 😉

        • Bensam123
        • 6 years ago

        Hmmm somehow I missed that section.

        Why are the numbers for MLC so low?

        That aside they make it seem like this stuff is ready for production, but I can only imagine how many years it’ll be before we see this stuff in devices.

      • Waco
      • 6 years ago

      EDIT FAIL

        • Celess
        • 6 years ago

        EDIT FAIL Combat (until article is updated)

        [code<]Characteristics NAND RRAM Comments Cell Size 5.44F2 4.28F2 provides better array size efficiency Tech Scale <25nm <5nm scales as is nanoparticle-based Program Perf 7MB/s 140MB/s performance is 20x faster Pg/Byte Erase Cant No Need not require Erase prior to Program Byte Program Cant Can greatly improves system performance Async Read .04MB/s 17MB/s provides faster NAND-like products Prog Engy/Cell 1360pJ 64pJ improves power consumption Cell Endure 3K cyc 1M cyc improves memory system lifetime Retention 1-3 y 20 y provides lifetime and reliability[/code<]

          • Waco
          • 6 years ago

          Thank you. 🙂 I was referring to my post as the fail though – I managed to create a new post when I tried to edit my last one.

    • ronch
    • 6 years ago

    Cool. But the real question is: will this tech actually find itself on the shelves?

      • Deanjo
      • 6 years ago

      Oh you pessimist. You doubted the flying cars too.

        • Grigory
        • 6 years ago

        They still suck, tho.

    • danny e.
    • 6 years ago

    sounds like it’s still several years off.

    also, cost?

      • bcronce
      • 6 years ago

      memresistors will be out in the next year, so this rram better get out soon, otherwise they won’t be able to charge a premium.

        • mesyn191
        • 6 years ago

        ReRAM has been out for years, its just been a mostly high cost boutique product.

        The “cheap” ReRAM is supposed be coming out sometime this year but I’d expect to still pay heaps of cash for GB’s of the stuff for quite a while.

        The supposed advantage with the stuff Crossbar says they have is that its going to be dirt cheap and fast right off the bat.

          • bcronce
          • 6 years ago

          Yeah, I’m not holding my breath for cheap or perfectly bug free for the first year or two, but there is a good potential for prices to come down fast as HP and Hynix have stated that they require very few changes and all simple to current NAND production to create their new memresistor parts.

    • fellix
    • 6 years ago

    What happened to the memristor technology. This is supposed to be the ultimate NAND replacement with virtually unlimited number of write cycles.

      • nanoflower
      • 6 years ago

      What happens with most fantastic sounding technology. They discovered some issue that made it less than fantastic. Either it’s too costly to produce in volume or it isn’t any better than existing products or perhaps even some legal issue came up. It’s seldom that a truly disruptive technology makes it from the lab to actually being a disruptive marketable product.

      • bcronce
      • 6 years ago

      Memresistors by Hynix, HP and Samsung are due out next year. HP said they had to push-back the roll out because they first wanted their NAND chips to clear out a bit more because NAND will be more expensive and slower. No point competing with themselves.

      Yes, these memresistors will have no write limits and will be used as system memory.

      edit: This isn’t just a “plan”, they have already spent billions re-tooling many of their plants for memresistors. They are financially committed.

      • stdRaichu
      • 6 years ago

      Hell, you could buy commercial [url=http://en.wikipedia.org/wiki/Mram<]MRAM[/url<] chips back in 2006 and I think ST-MRAM came out last year; this was stuff that was touted to replace DRAM back in the 90s. Like I suspected they might, they're being used for industrial/hostile environments and if you want even a GB of the stuff you're going to need to spend Christmas in the Nakatomi building. Whenever someone says "$new_storage_technology set to replace $existing_storage_technology!" you should mentally add "...perhaps maybe in five years if we're lucky, and probably more like a decade once the inconvenience of reality kicks in".

        • krazyredboy
        • 6 years ago

        I wouldn’t normally, say this, but I enjoyed too many statements, that reply.

    • chuckula
    • 6 years ago

    One [i<]million[/i<] cycles! Sounds very interestly as a replacement for NAND! As a replacement for RAM though, one million cycles ain't gonna cut it since you could likely wear out a chip in a few minutes a modern RAM speeds even if you have wear-leveling on the chip.

      • Hattig
      • 6 years ago

      Luckily it’s not a replacement for RAM, even though it apparent behaves like RAM as it allows true random access.

      The product promises a lot, but until someone puts a product in the market that uses it we cannot really ascertain how good it really is. Hopefully it will be available within a couple of years and disrupt the market (which seems to be stuck at 16GB on all mobile devices by default).

        • shaurz
        • 6 years ago

        Imagine 1TB on a phone, that would be mental!

          • lilbuddhaman
          • 6 years ago

          700gb of bloatware would take forever to remove.

            • Sahrin
            • 6 years ago

            Well the fastest NAND can saturate a 600 MB/s bus, so this stuff is theoretically 12 GB/s…you’d be OK, I think.

            • nanoflower
            • 6 years ago

            And your phone company would love you as you rack up the data rates filling up the phone.

            • Sahrin
            • 6 years ago

            You get charged for deleting things from local storage?

            • Diplomacy42
            • 6 years ago

            unlimited data @ 4g speeds is nice

            • Dizik
            • 6 years ago

            Yay for CyanogenMod, ParanoidAndroid, AOKP, etc. 😉

        • ET3D
        • 6 years ago

        Could be a RAM replacement for microcontrollers. 17MB/s random access speed would be enough for many applications.

Pin It on Pinterest

Share This