Comments Locked

44 Comments

Back to Article

  • SaolDan - Friday, February 3, 2017 - link

    Cool!!!
  • close - Friday, February 3, 2017 - link

    Hot!

    Is GDDR6 the successor of GDDR5 or GDDR5X? Seeing how the former is DDR and the latter is actually QDR I would have thought they'll also change the name.
  • Billy Tallis - Friday, February 3, 2017 - link

    They described GDDR6 as an evolution of GDDR5x.
  • close - Monday, February 13, 2017 - link

    GDDR5X is not DDR, it's QDR.
  • ddriver - Friday, February 3, 2017 - link

    Unfortunately reliability and endurance drop faster than the cost. I for one am not too keen on lower process node nand. I'd rather have them keep stacking it up, preferably in some ancient process, like 40 nm or so. Plenty of old fabs that can keep churning chips. Silicone itself is neither expressive nor rare.
  • ddriver - Friday, February 3, 2017 - link

    *silicon, not silicone lol
  • edzieba - Friday, February 3, 2017 - link

    What on earth happened to Hybrid Memory Cube? Intel and Micron were both partners on its development, but Micron basically doesn't even mention it anymore, and the only Intel product I can think of that is even slated to use it is Knights Landing.
  • Yojimbo - Friday, February 3, 2017 - link

    Last year at Hot Chips Micron included HMC in their presentations and someone from Micron called HBM a poor-man's HMC, or some words to those effects, if I recall.

    It looks to me like all TSV memory has penetrated the market much less than what people were previously saying. A couple of months ago I was wondering what happened to Wide IO, because I haven't heard anything about it. HBM has been used but only sparingly. I think it's just been too expensive to make thus far.
  • ZeDestructor - Friday, February 3, 2017 - link

    From the performance of Fury and P100, and comparing to 3DS DDR4 it looks like the interposer is the really hard part for HMC/HBM. Most likely Intel is keeping HMC for the top-tier stuff, the big Xeons on the Purley platform and Xeon Phi.
  • patrickjp93 - Friday, February 3, 2017 - link

    It's being used in enterprise. Intel uses it on Xeon Phi. There's a specialized variant of the Stratix X coming which uses it. Fujitsu makes Sparc CPUs attached to HMC. And IBM has a few specialized machines which use it too. It's just not a JEDEC standard, so it's not coming to consumerville any time soon.
  • wumpus - Saturday, February 4, 2017 - link

    You will never be able to add memory to an interposer, so I can't expect anybody to care about JEDEC standards in "consumerville". It might make it easier to build a product if there were competing memory suppliers, but I suspect that might be awhile.

    I just want to see CPU+HBM on an interposer (not necessarily 16GB, 4-8 would do) and the rest of it connected to 3d-xpoint (or similar, low latency-low cost) memory via "DDR[n]" like system (*not* PCI-e). Looks like that isn't happening anytime soon (no idea if Intel is blackballing potential AMD chips ready for such).

    My guess is that such systems would not only canibalize Micron's DRAM sales, but they would cut into any profit that crucial makes selling SSD systems. With the 3dxpoint acting as a massive cache/write-multiplaction buffer, you can add drives that are basically bulk silicon and let the system handle the flash issues (assuming that the drive handles the ECC). Of course if "the system" == "windows" then all bets are off.
  • fanofanand - Monday, February 6, 2017 - link

    "On the LED side, the four zones on the board are coming trolled through the bundled software."

    "I just want to see CPU+HBM on an interposer"

    HBM IS memory, so you are saying that what you want to see, you will never see. You appear to have great internal conflict about this.
  • fanofanand - Monday, February 6, 2017 - link

    copy & paste fail. Meant to show

    "You will never be able to add memory to an interposer,"

    "I just want to see CPU+HBM on an interposer"

    Sigh, I get why they haven't added an edit function because it forces people to think before they post, but it sure would be nice if we could edit within 10 seconds of posting or something, few people will post something for purposes of pure provocation and then instantly edit or delete it.
  • BrokenCrayons - Friday, February 3, 2017 - link

    I do hope we can move away from current NAND technology and onward to something a lot more durable. The idea of mainstream QLC SSDs with very low p/e endurance is absolutely nauseating to me.
  • TheinsanegamerN - Friday, February 3, 2017 - link

    " They are watching the demand quite closely and plan to be ready to introduce QLC NAND for certain cloud computing uses once the market is ready."

    Not sure where you are getting mainstream from that.
  • BrokenCrayons - Friday, February 3, 2017 - link

    I expressed concern over a potential future where QLC becomes mainstream rather than implying that future was a foregone conclusion. Admittedly, I probably should have stuck the words "if" and "before" in there someplace.
  • Kvaern1 - Friday, February 3, 2017 - link

    Just don't buy it then?

    You know, different users different needs.
  • fanofanand - Monday, February 6, 2017 - link

    I understood what you were saying BC, and I do agree. I think QLC has it's place, and that's in predominantly read-only functions.
  • romrunning - Friday, February 3, 2017 - link

    I second that. If people want lower prices, let them go TLC. QLC is a product that ultimately hurts SSD's reputation, with its low endurance.

    Micron, please focus on getting 3D XPoint SSDs into the market rather than the low-margin QLC junk.
  • patrickjp93 - Friday, February 3, 2017 - link

    If you have a database that's write little read often, then there's no need for super high endurance.
  • Yojimbo - Friday, February 3, 2017 - link

    Do you run into endurance problems? Or is it just a level of anxiety because you are imagining a ticking time bomb? Those pixels in your OLED phone screen are doing the same thing :|
  • BrokenCrayons - Friday, February 3, 2017 - link

    It's absolutely unproven anxiety over a looming, yet-to-happen failure. Then again, I have MLC SSDs in my two laptops at the moment (both 120GB Patriot Torches that have been very well-behaved). As for OLED, I don't own a cell phone.
  • lilmoe - Friday, February 3, 2017 - link

    If you're worried about endurance, you should really opt for higher capacity drives...
  • BrokenCrayons - Friday, February 3, 2017 - link

    They're old and yes, I'm kicking aroung the idea of picking up bigger drives eventually, but it's hard to rationalize spending any money at all on a Dell Vostro 1500 and Latitude e6400. They're both ancient.
  • bcronce - Saturday, February 4, 2017 - link

    If you had to choose between a 100GiB SSD with 300x writes or a 100TiB SSD with 3x writes, which would you want? Because that's where they're going. Turns out if you almost never plan on changing the data you've written, you can write a lot of data. Great for long term storage.
  • wumpus - Saturday, February 4, 2017 - link

    QLC gets 33% more space, not 1000x more space. Also Moore's law is struggling, so getting 100TB flash might be difficult (unless you are a datacenter who already has such space).

    A better question is how much wear an "SLC" mode has on such things: if they can confine the writes to a moving band of wear, they can leave the rarely written parts alone and still take the abuse of heavy writes.
  • MajGenRelativity - Friday, February 3, 2017 - link

    Why is GDDR6 necessary? Wouldn't HBM2 take over the high end market?
  • Mr Perfect - Friday, February 3, 2017 - link

    One would hope. I'm guessing GDDR6 will be for midrange cards?
  • MajGenRelativity - Friday, February 3, 2017 - link

    But if it isn't going to be high end, why bother spending even MORE money on top of GDDR5+GDDR5X?
  • fanofanand - Monday, February 6, 2017 - link

    I would love to see the industry move away from DDR period. It isn't a huge bump but it sickens me to know that every purchase of any product containing DDR, a few shekels get sent to the patent trolls at Rambus.
  • Yojimbo - Friday, February 3, 2017 - link

    Seems not. It sounds more and more like TSV memory will remain pretty expensive. Besides, I don't think graphics cards need that sort of bandwidth yet. GPU compute cards do, but not graphics cards. So for most graphics cards it ends up adding a whole lot of cost for some energy savings. Not worth it.
  • lilmoe - Friday, February 3, 2017 - link

    Cost...
  • vladx - Friday, February 3, 2017 - link

    ^This
  • wumpus - Saturday, February 4, 2017 - link

    Cost and expansion. Ideally I'd like HBM2 (supposedly lower cost than HBM) and DDR5/6 made out of 3dxpoint or other stuff that is cheaper than DRAM.

    But don't be surprised if phones stick to DDR[n] as well even though HBM should allow less power to be consumed (flagship models might loudly use HBM).
  • WulfTheSaxon - Friday, February 3, 2017 - link

    “In particular, they do not want to become a competitor to their valuable DRAM customers”? I think I see the slide where that comes from: “Partner as opposed to competitor to
    our customers” as a bullet-point under "Why Micron Wins". But did they forget that they sell enterprise and consumer SSDs and RAM, memory cards, thumb drives, and more both directly and through their Crucial and Lexar subsidiaries? All I can think they’re referring to is their spinning off MicronPC back in 2001, which seems like an odd thing to mention. >.<
  • Guspaz - Friday, February 3, 2017 - link

    I'm starting to think that 3D XPoint is going to be too slow to replace DRAM, and too expensive to replace NAND, even as a write cache (since SLC is already used as a high-endurance write cache for modern consumer SSDs). Unless the cost comes down by a huge margin, it's going to end up as a niche product that never makes it out of enterprise use where there's a need for absurdly high IOPS.
  • MrSpadge - Friday, February 3, 2017 - link

    Of course it won't replace DRAM in almost all usage scenarios. It's got a few orders of magnitude higher write endurance than NAND, which means it would still wear out pretty quickly if used as regular main memory.
  • wumpus - Saturday, February 4, 2017 - link

    It should be able to replace *some* DRAM, but not all. It looks like its best chance is to try to shove itself into the memory hierarchy. Something like:

    onchip cache->HBM dram (on interposer)->3dxpoint (DDR[n]-like bus)->bulk flash (pci-e/m.2 bus)

    The point about the "bulk flash" is that the 3dxpoint should take up most of the buffer duties and act like a cache for the flash drives. This should make SSDs more of a commodity. I'm guessing the biggest threat to 3dxpoint is 3dflash's* ability to get some deep wells to act more like SLC (especially in endurance) and act as it's own buffer circuit (this is the DRAM 3dxpoint "should" be replacing).

    * I'd think this works in 2d, but can't say I've heard of it being done.
  • fanofanand - Monday, February 6, 2017 - link

    Unless it is part of the package (whether that's on-die or within the chipset) then it is anti-consumer to try to wedge yet another product in the memory hierarchy. We need LESS points of failure, not more.
  • martinw - Friday, February 3, 2017 - link

    " Micron sees the memory market as a whole becoming less volatile"

    I see what you did there...
  • vladx - Friday, February 3, 2017 - link

    So 2017 will be the year where planar era finally ends.
  • Michael Bay - Saturday, February 4, 2017 - link

    They will still be making planar after you`re dead, just as "analog era" never ended.
  • Billy Tallis - Saturday, February 4, 2017 - link

    Micron should entirely cease making planar NAND within the next few years. By the end of this year, they will probably have migrated all of their own SSD product lines over to 3D NAND. At that point they'll only be making planar NAND to fulfill support/supply commitments for enterprise SSDs that won't reach EOL for another few years. I believe they've already transitioned to where you can't buy planar NAND from them without an existing supply contract.
  • Kristian Vättö - Sunday, February 5, 2017 - link

    For SSDs that is true, but for industrial, which is Micron's biggest cash cow, 10-15 year longevity programs are normal. They will continue planar for a good decade at least, albeit it will be a very small portion of total bit output.

Log in

Don't have an account? Sign up now