Comments Locked

54 Comments

Back to Article

  • Kakti - Wednesday, December 28, 2016 - link

    So this is what scares me about Optane aka 3d Xpoint...they seem on par or weaker than current NAND drives due to PCI-E lanes. The drives in this article say they are PC-E 3.0 x2 drives...that's a max of ~2gig/sec, already outclassed by Samsung's newest drives (and others?). Previously they showed I believe a 3.0 x4 drive that had similar numbers, somewhat below the top end NAND drives.

    Has anything been published that 3d-Xpoint aka Optane can use a ton of lanes or something? Five years ago the numbers these drives put out would have looked great, I don't think anyone expected Samsung to basically saturate the x4 line so fast.

    Unless Optane can basically run a ton of lines in parallel I think they might be beat for what the presumed asking price is. I hope Micron and Intel can pull something out of this tech!

    *disclosure:I own Micron stock and have lost my ass over the years (but improving this year!). Everything above is my opinion only.
  • patrickjp93 - Wednesday, December 28, 2016 - link

    Are you joking? Intel's Optane drives kick the snot out of existing SSDs. Look at the QD1/2/4 numbers for the 960Pro and Intel's Optane options. Intel is the clear winner by a mile and a half. No one gets close to QD32, no one. Intel's solution will feel faster even if the peak numbers don't look so hot on 1 and 2-die solutions.
  • Kakti - Wednesday, December 28, 2016 - link

    Yeah it's supposed to have much better access/seek times (in the 120-140k IOPs range?) but whether that translates to any meaningful increase for a regular user remains to be seen. It seems like IOPS really matter on big databases and not necessarily a regular PC or most workstations.

    I dunno, I just feel this is too little to late with the small size of the new "cache/RAM drive" compared to just rolling a straight top end m.2 drive and 32gb+ of RAM.

    Anyone can make a 50gb+ ramdisk already on their workstation for a couple hundred bucks if they really need it.

    So you get 32gb of RAM and either a PCI-E x4 NVME drive or an m.2 x4 NVME drive with 500gb+...where does that leave this octane drive?

    High end users have too much ram/fast SSD to cover it. On the low end (4gb-8gb ram) it would probably shine the most but I don't know if they'll be offered in this class of products It seems like servers and databases these will do very well, so that may be the main market. I'm very curious to see some very detailed *cough* reviews once these hit the street.
  • beginner99 - Wednesday, December 28, 2016 - link

    3D XPoint is more like RAM so yes it's random performance (access times) will beat any SSD easily. And let's not forget that it was exactly this (latency) that made the biggest difference from SSD to HDD and not the sequential read/write speeds. So I can easily see that a 16 GB optane cache can make a system even with SSD much more responsive.

    And lets not forget that compared to a RAM-disk it is non-volatile. So all the cached OS files and applications are there so you will probably get a CPU limited boot and application load times! It sounds pretty cool on paper.

    Realistically even for an enthusiasts it will almost certainly not be worth it if you don't run high access databases or do some other niche level stuff.

    I'm somewhat undecided about price. The solution must obviously be cheaper than just going with an SSD. On the other hand I fear it will always be cheaper to get a faster SSD instead of adding a cache drive. Eg. if the 32 gb cache drive cost about the same as the difference between a average SSD and a 960 pro, I would choose the later,
  • iwod - Wednesday, December 28, 2016 - link

    But meaningful performance difference? SSD, at latency and certain I/O is 1000x faster then HDD. The same cant be said about Xpoint and SSD. And we surely reach a point where performance isn't in the I/O.

    i.E Starting your App from RamDisk isn't that much faster then SSD.

    What i really want to see though, is QLC, 3D stacked NAND as Storage and Xpoint as main drive. Even a 64GB main drive is good enough for me.
  • ddriver - Wednesday, December 28, 2016 - link

    Until I see optane in action I call BS. So far it looks like it will mostly have much better caching and much larger cache buffers, which will be enough to improve the IOPS regardless of the actual bulk storage medium..

    They claim "bit addressable" for optane, but that's BS too, even RAM is NOT bit addressable, not even CPU registers. RAM is BYTE addressable. Bit addressable for real would be a fruitless endeavor because the amount of extra traces will not be worth the returns, as nobody bothers with bit access.

    My bet is the medium will not even be byte addressable, they will just add that in software via bitwise ops. Much like the IOPS boost will likely be due to caching and only for cached data. All this is perfectly possible to do with NAND flash as well, as usually the industry is holding back on very logical and straightforward improvements in order to milk the most out of everything.

    So far optane has been rather hypetane... and how we get news of it making it into actual products. Oh wow, 16 gigs of optane, that's almost enough to install windoze. Paradigm shifting!
  • mdriftmeyer - Wednesday, December 28, 2016 - link

    Stop it! You're making too much sense. What you describe is a dagger into this 3D crap from Intel. Put in that 1TB or 2TB PCI-E based NVME drive and save money in the process while your system flies.
  • Billy Tallis - Wednesday, December 28, 2016 - link

    The PCIe x2 bit comes only from the months-old leaked roadmap, and we have no official confirmation of that. And for what it's worth, those roadmaps showed that only this first generation of cache drives would be so limited, while the next version will move to PCIe 3 x4.

    But the biggest advantage of 3D XPoint memory is low latency, which doesn't require a lot of PCIe lanes. And with only one or two dies, the bandwidth limitation might be between the controller and the memory, rather between the controller and the host. Larger drives suitable for primary storage rather than just caching will have the same parallelism advantages that flash SSDs enjoy, and should be much better able to deliver high throughput.

    In the meantime, these cache devices will probably offer sufficient throughput given their small capacity and the low performance of the hard drive backing devices, while not wasting power on more PCIe lanes than needed to offer the most performance benefit possible without abandoning the mechanical hard drives entirely.
  • drajitshnew - Thursday, December 29, 2016 - link

    My laptop is i7-4700mq lenovo y510p. It has a m2-2242 slot. Basically, ruling out any high performance or mainstream drives.
    My mom's laptop is i5-4200u, costed half as much, weights half as much and has an 2280 slot.
    I anticipate lenovo will DELIBERATELY cripple the sku's with optane to X2--2242 designs.
  • unityole - Wednesday, December 28, 2016 - link

    people like you only look at sequential eh? know why SSD is so much faster than HDD? its because of latency hence the QD1 4k random read/write performance. optane will bring QD1 4k random read/write so much closer to that speed of ram, samsung isnt even able to compete. the sequential speed you see won't be used as often as you think brother.

    time to move on and really learn what tech is about and what benefits and what doesn't.
  • vladx - Wednesday, December 28, 2016 - link

    True but like Kakti already mentioned above it will have little to no impact in consumer space. So good for enterprise, meh for the rest.
  • beginner99 - Wednesday, December 28, 2016 - link

    I disagree. The advantage of even early SSD over HDD was to reduce stutters / waiting time on the disc to basically 0. Consumer space is mostly QD1 and hence latency matters a lot. It will make lag/stutters disappear completely.
  • Nagorak - Wednesday, December 28, 2016 - link

    The thing is SSDs are already so fast that most users will simply not notice a difference. I don't have any stutters or slow access issues currently with my SSDs and they're only budget Samsung models. It's a case of diminishing returns.

    It's kind of like the difference between VHS, DVD and Blu-ray. There was a huge difference in quality in portability in moving to DVD, so it caught on strongly. The move to Blu-ray was much more subtle and a lot of consumers response was "meh".
  • Shadowmaster625 - Wednesday, December 28, 2016 - link

    Tis true. I have a dozen systems with old Intel X25-M drives. And I have a few systems that have newer SSDs like Samsung 850 pro and I even have a newer Intel SSD with a skull on it. There is no noticeable difference between any of these. As long as they have an SSD they are good.
  • bcronce - Thursday, December 29, 2016 - link

    SSDs are literally magnitudes faster than mech drives, but even my wife's Samsung 250 Pro M.2 can only manage about 100-200MiB/s for any single-threaded operations because of QD1 speed. It is nice that I can have many things going on at the same time and still get 100-200MiB/s, but it would be nice to be bandwidth or CPU limited and not IO latency limited.
  • TheinsanegamerN - Wednesday, December 28, 2016 - link

    For general consumers, SSDs already eliminated lag/stutter. Non professionals just dont have a use for a drive this fast.
  • lobz - Wednesday, December 28, 2016 - link

    I beg you to try to find any stutters starting and running consumer applications on my PC. It's sporting a 950 PRO and a 4790K, so nothing special or extraordinary. Still, I challenge you to find _any_ noticeable stutters and/or waiting time.
  • StrangerGuy - Wednesday, December 28, 2016 - link

    Stuff you mentioned are already indistinguishable on a SATA3 SDD vs a RAMdisk. No flash is faster than DDR3/4.

    Another symptom of why consumer PC tech is so bloody boring these days, besides GPUs everything else new is hitting massive diminishing returns in real world usage.
  • doggface - Wednesday, December 28, 2016 - link

    I am willing to admit I may be wrong, but I think anytime you improve random performance their will be a noticeable increase in responsiveness.

    However, having used RST before, I must admit to some scepticism here. I would rather have a 500gb Sata SSD, than a 16Gb OPTANE + 500GB spinning Rust.

    Mainly the thought of a hdd in a laptop makes me cringe.
  • Anato - Saturday, December 31, 2016 - link

    Agree. RST limit for 64GB is arbitrary, old and stupid. 16GB is too low for meaningful write-back cache, I would prefer 16GB X-point + 256GB SSD under 1 controller on m2-2280 and 2TB HDD. Or only 256GB SSD as cache to 2TB HDD.

    Second thing is, I'd like to see TLC NAND made to perform as SLC and MLC depending on how much data is actually stored in drive. This could be change on the fly as from only SLC -> partial MLC -> partial TLC. I don't mean implementing this would be easy, but there are much tougher problems in SSD than this.
  • bananaforscale - Friday, December 30, 2016 - link

    For the time being, yes. Wait for a while.
  • Response - Wednesday, December 28, 2016 - link

    16 GB? Is this a joke?

    There is NVDIMM, which is just DRAM with backup of NAND, it has the full speed of DRAM and endurance. Other technologies are needed for large capacities 500GB and more. 16GB we have NVDIMM which is much better and faster.
  • Billy Tallis - Wednesday, December 28, 2016 - link

    The platform changes (both hardware and software) necessary to support Optane NVDIMMs are massive compared to what it takes to support Optane NVMe SSDs. Optane NVDIMMs are coming to the enterprise market and maybe eventually to the consumer market, but first 3D XPoint will have to prove its value in NVMe products that can be brought to market without requiring an overhaul of key CPU subsystems.
  • patrickjp93 - Wednesday, December 28, 2016 - link

    It has roughly 1/4 the speed of RAM when truly swamped. Don't kid yourself,
  • mikegrok - Wednesday, December 28, 2016 - link

    This is awful.

    The software stack for nvme takes 1/20th time to execute for a single transaction compared to ahci. This is where much of the speed boost comes from, to say nothing of the speed boost from having parallel threads of disk transactions managed in hardware instead of juggled through software buffers for the single command queue in ahci.

    What happens if you have a bad windows update that kills the boot drive, such that you need to do an OS reinstall, or have to repair some ntfs corruption? When you boot off of a backup disk, good luck getting the cache to associate with correct ssd, instead of the boot disk. What happens if you are attempting to use whole disk encryption?

    There is a use for optane, but I don't think this is it. Allocate some of the optane to hold the write lock state for a file server, or the row lock state, or hold the current write transaction for a database. optane can commit single word write transactions, without needing to commit a 4k block. optane's sequential speed is equivalent to a samsung 960, but it can commit a one word block much quicker, effectively greatly speeding up a database server.

    I thought that optane could have been a good local backup to RoCE, which provides a 1.2 microsecond transaction time for small transactions. Mellanox is a good keyword when researching RoCE.

    -Michael McMillan
  • haukionkannel - Wednesday, December 28, 2016 - link

    That is why this is first committed to enterprice use. This is minor "customer" class version for optane. The big companies use it differently in data Centers.
  • hahmed330 - Wednesday, December 28, 2016 - link

    Hmmm... I wonder if these dies are giant 600mm^2 two-layer 128Gb (16GB) die that's what I heard from others then colour me not impressed nor for the speed since 2GB/s for a cache is very unimpressive one would be better of using proven and tested pure SLC caches.
  • witeken - Wednesday, December 28, 2016 - link

    They're 227mm².

    http://www.anandtech.com/show/9470/intel-and-micro...
  • jjj - Wednesday, December 28, 2016 - link

    Another misleading teaser to keep the hype up.They tease it now, they show it at Computex and then maybe it ships at the end of the year lol.
  • ex_User - Wednesday, December 28, 2016 - link

    Yeah, coming SOON™...
  • Morawka - Wednesday, December 28, 2016 - link

    wow 2 layers, these are super simple.. someone could make these in their lab with little trouble.
  • sharath.naik - Wednesday, December 28, 2016 - link

    What is the point of this? Does not drives already have rams built in with capacitors to handle data loss. Why go with a more expensive solution? I do not thing capacitors cost that much when used with a RAM. If this was used as an SSD then yes it is a logical replacement for nand SSD, but this style of usage seems pointless.
  • Billy Tallis - Wednesday, December 28, 2016 - link

    Consumer SSDs don't have power loss protection. They rely on filesystems to be reasonably crash-tolerant, which is fine for the consumer market, especially when most machines are notebooks with built-in batteries and will cleanly shutdown or hibernate before losing power.

    3D XPoint memory is supposed to be cheaper than DRAM, and thus also cheaper than DRAM+supercaps. But I don't think this is primarily about ensuring data integrity. It's about improving performance beyond that of hard drives, but without raising the overall price per GB up to the level of NAND flash SSDs.
  • Holliday75 - Wednesday, December 28, 2016 - link

    Can this tech replace the cache on RAID controllers? No need for a battery that loves to fail. Power lost during write process. No big deal. Next power up data written as its stored in non-volatile memory.
  • Billy Tallis - Wednesday, December 28, 2016 - link

    I'm not sure if it's quite fast enough for large RAID controllers, and it's not clear whether endurance would be an issue. We don't really know what the per-die raw performance is, and only have hints about what kind of performance can be achieved through a SSD controller.

    RAID write caches for the sake of data integrity rather than performance don't need to be very large, so things like MRAM are also an option.
  • Holliday75 - Thursday, December 29, 2016 - link

    Good points. Thought just popped in my head as I used to work in a data center and hated getting tickets for failed batteries. Had to take the darn server offline just to swap a battery. The batteries used on HP raid controllers liked bulging and failing.
  • Magichands8 - Wednesday, December 28, 2016 - link

    My God, the turd has finally dropped. The mystical and vaporous Optane begins to appear. But WTF is with this 16GB nonsense? Codenamed Stony Beach, eh? It's definitely been a long and rocky road getting it here but what's the Point at 16GB? You could have the fastest memory and interface ever devised but if you release it at 64KB it's worthless. Guess they need some guinea pigs to cut their teeth on. By the time this is out Samsung et. al. should have 10TB NAND SSDs on the market running over PCIE. Of course, they should have had that out 2 years ago and we probably won't actually see it until nigh on 2040 but either way Intel better get its ass in gear.
  • Murloc - Thursday, December 29, 2016 - link

    I suggest you read about memory hierarchy because your thinking is off. There's nothing wrong with it being 16 GB if it really is fast.

    And L1 cache in CPUs is very small so your statement about the fastest memory being useless if it's small is simply wrong.
  • Magichands8 - Thursday, December 29, 2016 - link

    Unless of course you have more than 16GB to move. Which was the point. And while it's fascinating that you want to randomly talk about CPU caches that's only another reminder that you didn't get my point.
  • boeush - Wednesday, December 28, 2016 - link

    I can see only one potential application for n consumer space, at least as long as capacity remains low and cost per GB high.

    And that application would be in mobile devices to enable fast hybernate/resume with almost zero power use while hybernating - as compared to any sleep state that still has to power the RAM...
  • Kakti - Wednesday, December 28, 2016 - link

    Phones may actually be a pretty good scenario for 3D-Xpoint/Optane replacement.

    Along a similar vein then are those super cheapo netbooks/chromebooks that still use eMMC for storage. I'm hoping this year we finally see the end of those god awful 16gb and 32gb eMMC drives. But then again NAND may just end up replacing eMMC for these low drives.

    At the end of the day, we really need more information on how big/dense these things will be. Are we going to continue seeing DIMM-esque drives and modules that run in the dozens of MB each, or will we be seeing multi-terabyte size products in the consumer space? Are these always going to play the role of "smaller, faster cache" in a RAM > Optane > NAND design, or do they plan (hope) that this will replace NAND eventually?

    Moving offtopic for a second, in regards to DIMMS vs NAND, is it something inherent in the physical design and architecture of DIMM Memory that makes it so much less dense than comparable NAND? Why does a given size of a DIMM only have say 8gb, 16gb or even "crazy" 128gb of RAM, when an M.2 drive can have 512gb-1tb? I assume they're on same or at least very similar lithography so assume it's just a simple design difference that makes RAM so much more volume hungry?
  • extide - Thursday, December 29, 2016 - link

    eMMC is just a protocol -- eMMC drives still use NAND flash just like sd cards and everything else
  • extide - Thursday, December 29, 2016 - link

    With DRAM you basically have a capacitor for each bit, and whether it's charged or not is how it knows if it's a 1 or a 0, but the charge in that capacitor drains out very quickly, which is what makes it volatile and that's why you need to keep power because it has to constantly read the cell and then refresh the charge in the cell. Requiring a cap means that it's hard to make the cell sizes small. For Example: a single die of modern top end latest lithography DRAM has about 8 or maybe 16Gbit capacity. NAND can store 384Gbit in the ~same space right now with IMFlash's 3D NAND.
  • Kakti - Friday, December 30, 2016 - link

    Gotcha, thank you for the reply extide.
  • tuxRoller - Wednesday, December 28, 2016 - link

    This explains the new instructions and platform support for pmem:

    https://lwn.net/Articles/674752/
  • Billy Tallis - Wednesday, December 28, 2016 - link

    That only applies to NVDIMMs, not PCIe NVMe devices even if they have Optane memory on board. Also, it's out of date: https://software.intel.com/en-us/blogs/2016/09/12/...
  • tuxRoller - Thursday, December 29, 2016 - link

    The nvdimm form factor is irrelevant (lightnvm works similarly). This is speaking about the challenges of pmem, and yes, it's a bit old but it actually provides some details as to why Intel is requiring a platform to support optane.
    http://pmem.io/
    Also, the clwb command was relegated to the SDK in the October update.
  • Billy Tallis - Thursday, December 29, 2016 - link

    LightNVM has approximately nothing to do with NVDIMMs. LightNVM is about moving the flash translation layer to the host CPU but otherwise still operating as a block device rather than memory mapped.
  • tuxRoller - Friday, December 30, 2016 - link

    Oh boy, this will be fun.

    https://en.m.wikipedia.org/wiki/NVDIMM

    I mentioned Lightnvm BECAUSE those pmem libraries are useful where the host is responsible for ensuring data persistence.
    Also, the dimm ff was, tmk, the first available and had been used for os support bringup wrt pmem.
    This is about creating a "native" solution to that problem instead of using dax. Obviously there are still some folks left to convince.
  • Byte - Wednesday, December 28, 2016 - link

    Wooo moar level of cache!

    L2, L3, L4/eDram, RAM, Optane, SSD RAM buffer, SLC Buffer, TLC/MLC Nand
  • Xajel - Thursday, December 29, 2016 - link

    So, Can we use it also as Cache to fast NVMe SSD ? looking that 3dxp is using PCIe x2 compared to the newer NVMe SSD's which uses PCIe x4, the NVMe will have higher bandwidth than 3dxp but the later will have lower latency...

    So it might be good for caching SATA SSD's, but I don't think these will be good enough for NVMe SSD's. at least not in this very first gen. of 3dxp...
  • BrokenCrayons - Thursday, December 29, 2016 - link

    I'm not really interested in Optane from a caching perspective, but my interest is moreso in its high endurance. Skip that cache silliness and figure out a way to lower the price per gigabyte to the point where I can dump 3K-5K write endurance (or less) on existing solid state drives in favor of SOMETHING that can shrug off that number as insignificant.
  • Meteor2 - Friday, December 30, 2016 - link

    Quite simply, I'm struggling to see the problem XPoint solves. NVMe is fast enough for persistent storage, is it not? Even for big, busy databases.

    Another way of thinking about it: what new thing will this enable? What will we be able to do that we couldn't do before?
  • BrokenCrayons - Friday, December 30, 2016 - link

    Do you remember Intel's Turbo Memory? It was a 512MB to 4GB mini PCIe card designed to improve storage subsystem performance on Santa Rosa-based systems. Anand reviewed it and two articles ended up on AT in 2007. Then there was small SSD caching on more recent Core i-whatever systems using small solid state drives in front of hard drives. XPoint is seemingly yet another iteration of this concept of sticking something between the storage and RAM to speed things up. I'm not saying its a bad idea (at least not until benchmarks come out that can measure its benefits or lack thereof) but it seems like Intel's still iterating on the idea.

Log in

Don't have an account? Sign up now