back to article No NAND's land: Flash will NOT take over the data centre

NAND is not taking over the data centre. Estimates of shipped disk and SSD capacity out to 2018 shows that, while flash will continue to grow faster, disk capacity will still outstrip that of flash by eight to one four years from now. SSD_vs_HDD_ships_To_2018 Total SSD capacity shipped is falling more and more behind disk …

  1. toughluck

    "Each shrink in NAND geometry seems to require costlier manufacturing processes and more over-provisioning to keep endurance, expressed as drive writes/day for five years, up at acceptable levels."

    And shingling, TDMR, HAMR cost nothing to implement?

    Between 2000 and 2010, over some 10 years, mainstream hard disks went from ~40 GB to 2 TB, a 50-fold improvement without increasing the price too much (allowing for the overblown Thai flood hype). That's an average increase of nearly 50% per year. If this pace kept up, mainstream disks now would be almost 10 TB in capacity, and yet mainstream is still at 2 TB, very slowly moving towards 3 TB, never mind 4.

    The recent breakthroughs that will allow more than 4 TB are ridiculously expensive and it seems that disk cannot break through this ceiling.

    Going by the same chart, I see that SSDs are predicted to grow by almost 2000%, while HDDs only by 123%.

    However, assuming the data for two last years and estimate for this one are accurate, this is an interesting extrapolation.

    It predicts that HDD growth is expected to increase or keep at a steady rate, while SSD growth is supposed taper off, astonishingly so -- it grew by 120% in 2013, then 85% this year, and they are expecting this trend to continue and growth to decline further, while HDDs are not going to be affected at all? I call bullshit.

    Oh, and they've got SSD endurance wrong. Taking a 480 GB SATA 3 drive at maximum speed (600 MB/s=52 TB/day) and 10000 writes/block, assuming you never stop and you never read this data, you get 92 days of useful life. That's still extremely high endurance. Lower this by a factor of 10, to 5.2 TB/day, and you get almost three years of useful life. However, 5 TB/day on a 480 GB drive? Who writes (and overwrites) this amount of data daily? If there's a usage pattern that fits this requirement, I suppose the user is getting paid well more than enough to cover disk replacement.

    1. Alan Brown Silver badge

      "Between 2000 and 2010, over some 10 years, mainstream hard disks went from ~40 GB to 2 TB, a 50-fold improvement without increasing the price too much (allowing for the overblown Thai flood hype)."

      Between 1994 and 2000 there was a steady increase to 100Gb, then there was a 4 year pause before they romped up to 2Tb and only slowly increased from there. At the same time, disk reliability has plummeted and the drive makers know it or they wouldn't be busy shortening warranties to 12 months (which is arguably illegal for warranties offered to endusers in the EU)

      In the meantime, flash capacity keeps increasing, flash reliability keeps increasing and flash warranties keep extending. 10 year warranty on the Samsung 850Pro may seem extreme but it's likely where the flash market is going.

      Now that 3D manufacture is mainstream, makers don't need to fight with tiny geometries and that should see an explosion of capacity at any given price point over the next 12-18 months - assuming the demand from phones doesn't eat the lot.

      HAMR has promise. Shingling is turning into a clusterfuck - the drives are so sensitive to vibration that they'll never work reliably in arrays.

    2. Anonymous Coward
      Anonymous Coward

      And shingling, TDMR, HAMR cost nothing to implement?

      Comparatively speaking, yes, they do. Shingling is free except for a bit of R&D time (already spent by the companies), TDMR requires multiple front end readers and relatively smaller tweaks to the digital back ends in the SoC, two extra lines on the flex cable, and another channel in the preamp (or the ability to repurpose the already large number of readers in a preamp).

      HAMR requires adding a laser to the writer, and timing information coming from the SoC.

      So TDMR/HAMR add a very small amount to the BoM ( less than $3 I'd guess) but really require almost nothing in the way of capital infrastructure like building a fab. The spinning rust companies can keep all their writers, testers, etc the same, but it will cost them a bit of R&D to get everything developed. If I had to guess, the switch to TDMR will be nearly expense free, HAMR will take a bit more (laser reliability is horrid from what I've heard), but nothing looks as hard as getting folks to risk more than $3B on a new fab.

      From what I've heard the flash guys are very reluctant to go off on mad capex runs due to the risks of overbuilding. They're pretty happy to milk what they've got going now.

    3. Gary Bickford

      toughluck is right. The scientific progress of 2000-2010 was almost entirely the output of one facility, Seagate's research facility in Pittsburgh. That was shut down in 2008, and all of the PhDs were fired (rather callously in fact). These were the world's experts on the solid state physics involved. None of them stayed in the industry.

      The technologies mentioned in the article are all derived from that work, which is no longer going on. We are just seeing the productization now of that research. Unless some company decides to invest $1B or so on a new advanced research facility, and finds a new set of world's experts, and gives them five to 10 years to build a new science, there will be no (zero, nada, nothing) new advances in spinning rust after the ones mentioned. Since Seagate and WD between them account for well over 90% of the total manufacturing capacity, that is unlikely to occur.

      IOW, HD capacity, performance and cost/performance will probably stall out by 2016.

      Also, industry estimates show that the cost/capacity curve of SSDs and HDs will cross in late 2016. I've seen the analysis. Newer SSDs have a higher lifetime than HDs, for almost all applications including the huge server farms that employ the majority of drives. So HDs are going to be a fading product from 2016 onwards.

      1. Anonymous Coward
        Anonymous Coward

        I'd love to see any analysis that shows $/GB for SSDs matching that of hard drives in two years time.

        Maybe it does if you extrapolate from the past five years, but if so it totally ignores economics (you know, supply and demand) because Chris is right, the capacity exists to make only a few percent of the total TBs of hard drive output in the form of flash. Even if you assume a very optimistic 10x growth in flash bits and demand growing at 0% over the next two years, there isn't nearly enough flash to go around! The result, the steep drop in $/GB flash prices is going to hit a floor far enough above $/GB hard drive prices to compensate for the performance difference, as economics says it should.

        Such a floor would add some certainty for the investment required to build additional flash fabs, but we're still talking a minimum of a decade before enough capacity could come online to replace hard drives. Probably longer given how insatiable the world's storage appetite is. Luckily, there is little reason to replace hard drives in a lot of roles that consume a lot of hard drives. For archival backup where tape is too slow, or disk to tape backups, flash gains you nothing. Ditto for applications that require a lot of capacity and sequential I/O - flash gains you nothing over striping a bunch of hard drives together.

        Much storage isn't performance sensitive to the point that all flash offers any advantage over a two tiered approach with 5-10% flash for the hot spots and 90-95% hard drives for the rest. I will agree that all laptops/desktops shipping in 2-3 years will be all flash, simply because the per bare drive minimum price is lower for flash than hard drives, so once the amount of flash included in those minimal drives is sufficient for a $250 PC (let's say 250 GB) that's what it will include.

        It remains to be seen how long the strings of 3D NAND can be made, but the most optimistic projections I'm aware of show it petering out somewhere between 128 and 256 cells. We're already at 32, so there's not much more to look forward to there unless they figure out a way to extend that. If they don't, NAND $/GB hits a wall, and we better hope for a competing technology like memristors to pan out!

        1. Anonymous Coward
          Anonymous Coward

          But it's not just speed where flash has an edge. There's size and longevity to consider, too. Both can mean firms may be willing to take a premium for either more room or a lower TCO (which one would think would be a more important metric than $/GB).

          "Even if you assume a very optimistic 10x growth in flash bits and demand growing at 0% over the next two years,"

          Do you take into account that one firm is already in mass production, a second is on the verge, and the remaining two (and we're talking major chip firms) are due in 24 months if not 18?

  2. Pavlov's obedient mutt

    because ...

    what comes after is breathing heavily down the neck of flash

    why invest billions on a technology that's already approaching it's sunset period.

    No, no no - flash isn't dead.. of course not, but think in terms of 3~5 years time

    I would like to have seen a chart like this with solid-state vs spinning - specifically leaving out flash as the media

  3. Lusty

    Ah but

    How much of the shipped disk capacity is RAID. Drives over 2TB shouldn't use RAID 5 or 10 for recoverability reasons so there has been a massive uptick in capacity sold just to cover this excess.

    Similarly in the home scenario, quite a lot of people have recently started using RAID to protect their family photos in Synology type devices so there has been an explosion in shipped capacity.

    Fast forward a year or two and people will start using flash for capacity at home as well as using RAID to protect that flash because it will be cheap enough to do so. The data centre will be a similar story as data stops being tiered to flash and starts being stored on flash.

    This data set is also extremely skewed by Google, Facebook and Microsoft who have been and continue to build out their cloud infrastructure. Zuck is on record as saying he'll buy flash that has a write life of a handful of writes if the capacity is large enough instead of spinning disk just for the power savings and latency improvement on the image store.

    Extrapolating figures based on current shipping volumes is a dangerous game, never more so than when nobody actually wants the old technology.

  4. AnoniMouse

    The writing is on the wall

    >> The huge great problem is $/GB. New disk technologies such as shingling,

    >> TDMR and HAMR are upping areal density per platter and bringing down

    >> cost/GB faster than NAND technology can.

    Disc technologies are heading towards their last gasp. HMR and TDMR have tahen FAR longer to bring to market than predcted. They are not just shrinks, but new technologies; and no follow-ons are apparent to follow in their wake. The fundamental limitations are a) domain size; and b) discs are 2 dimensional.

    3D flash on the other hand is only just getting started. Already stacking 100 layers is believed to be feasible. That's equivalent to 11 years growth in capacity at 40% CAGR. Noone is envisaging HDD technologies continuing to improve at anything like that rate.

  5. Alan Brown Silver badge

    "what comes after is breathing heavily down the neck of flash"

    When it arrives and is selling in quantities in the shops, THEN it's breathing heavily down the neck of flash.

    In the meantime it's still another promising solid state technology - we've seen plenty of those over the years, most of which simply haven't panned out.

    Once XYZ alternative technology arrives and is competitive, you can bet your booties that fabs will start producing it. Noone's betting the farm on NAND

    This isn't a contest between "flash" and "other sold state tech", it's a contest between "spinning rust" and "solid state that's good enough to do the job better than spinning rust"

    It's arguable that flash is already cheaper than spinners, once full lifecycle costs are taken into account (this includes power consumption and the labour costs associated with replacing failed devices as well as those associated with shorter/longer wait periods for data.)

    Our experience so far is that business-grade SSDs simply DO NOT FAIL in average service life (6-8 years here) vs 10% failure rate of HDDs - which has necessitated RAID1 on hdds for the last decade if you don't want staff downtime because of a dead PC.

    1. This post has been deleted by its author

      1. dkjd

        Re: "what comes after is breathing heavily down the neck of flash"

        to quote from the cited paper "physically isolated through the segregation of inter-granular non-magnetic Cr-rich [5]and oxide materials [6]".

        oxide materials = rust in my book

        and it spins!

        1. jabuzz

          Re: "what comes after is breathing heavily down the neck of flash"

          Only iron oxide is rust, all other metal oxides are just that oxides. So given the platters are either glass or aluminium, that the recording layer is devoid of any iron based materials there is no "spinning rust" in a hard drive and it has been that way for many years.

          http://en.wikipedia.org/wiki/Rust

      2. phuzz Silver badge

        Re: "what comes after is breathing heavily down the neck of flash"

        Every time someone complains about "spinning rust" it makes me want to use the term more.

        Old names linger on, after all, we still call it a Solid State Drive despite the lack of moving parts.

      3. Gary Bickford

        Re: "what comes after is breathing heavily down the neck of flash"

        I know someone who spent eight or ten years in the industry, at the research management level. They still call it spinning rust. I think it's kind of an insider thing.

        1. pyite

          Re: "what comes after is breathing heavily down the neck of flash"

          In my experience, people use "spinning rust" in the same condescending and passive-aggressive way as "legacy."

        2. Charles 9

          Re: "what comes after is breathing heavily down the neck of flash"

          I think the term you're looking for is "meme". The term has ascended beyond its clinical definition, much as "xerox" and (as mentioned above) "drive" have become memes. Who cares if they're not exactly right? They still evoke an appropriate image, just as the icon of a floppy disk still evokes the image of saving, so we're gonna use it regardless.

      4. Anonymous Coward
        Anonymous Coward

        Re: "what comes after is breathing heavily down the neck of flash"

        @Symon:

        Spinning rust times infinity, jinx, no comebacks.

  6. Trevor_Pott Gold badge

    Chris: while I agree that most individual drives shipped into the datacenter will be of the spinning rust variety, I believe the missing element here is hybrid setups. "All-flash" may be a bit of a rare beast for some time to come. We flat out don't have the global fab capacity to replace rust disks, even if we were willing to.

    But by the same token, rust disks are inadequate to the job. The immediate future is, IMHO, flash/rust hybrid setups. In array form or hyperconverged form.

    We'll see about "all-flash" for primary storage in a few years. That said, I'm working with a few startups that are making damned fine data intelligence/tiering software. I suspect "all flash" for primary workloads is 3-5 years away, with hybrid for tier-2 apps and "all rust" (or "cloud") for archival.

    It's all about the Benjamins, and the squeeze is on for the rust vendors. They are being moved into the commodity world, inch by inch. 15K SAS? Somehow, I don't think we'll be needing that for much longer.

  7. Michael H.F. Wilkinson Silver badge
    Coat

    If it is a disk world .....

    DOES THE PERMANENT ERASE FUNCTION SPEAK IN CAPITALS ONLY?

    Sorry, couldn't resist. Mine's the one with "Thud" in the pocket

  8. Fazal Majid

    Last hurrah

    You've got to wonder how much of this is driven by the transition from tape to disk. Certainly primary storage on laptops and enterprise first-tier is going all-flash, and mobile was always thus. The HDD iPod was discontinued. Hard drives seem to be increasingly relegated to cold storage.

    1. chris 17 Silver badge

      Re: Last hurrah

      Not sure how well a cold disk will respond after a few years on the shelf.

      Use it or Loose it

      1. Charles 9

        Re: Last hurrah

        "Not sure how well a cold disk will respond after a few years on the shelf."

        Compared to flash, I hear it stores better. Meanwhile, tape is only economical these days for enterprises. For the consumer market, it's pretty much hard drives or bust for the time being. To that end, I double-provision with a one-year rotation and use parity archives within for the occasional bit rot.

  9. This post has been deleted by its author

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like