Intel Optane SSD DC P4800X opens new frontiers in datacenter storage

Intel’s high-performance 3D Xpoint (say crosspoint) memory technology promised far-reaching changes for storage, memory, and computer architecture when the company first took the wraps off the technology in mid-2015. Of course, big promises begot high expectations, and despite a hunger for more information, we’ve heard relatively little about how 3D Xpoint will perform in the intervening time. That all changes today. 3D Xpoint will power Intel’s first Optane device, the SSD DC P4800X, and that device demonstrates the wide-ranging potential of Optane to change how we think about the balance between memory and storage in a system.

Optane from the ground up

The SSD DC P4800X launching today packs 375GB of 3D Xpoint storage on a PCIe add-in card form factor. Intel is keeping the physical details of its Xpoint media under wraps, but we do know some of the basics of the stuff now. The first Xpoint dies are being fabricated on a 20-nanometer process and can hold 128 Gbits each. Intel says that because the Xpoint medium itself relies on low-resistance and low-capacitance interconnects in the back end of the device, it can switch a thousand times faster than NAND. Consequently, Xpoint delivers much lower access times and much higher performance.

3D Xpoint also offers much finer access granularity than NAND. Instead of working on pages with many kilobytes of data as NAND does, Intel says the structure of the Xpoint device allows for bit-level access. Practically, that means Xpoint can work with words of data, so it’s a natural fit for delivering cache lines like DRAM does. Even though it’s not as fast as DRAM, Xpoint can be manufactured with 10 times the density of that memory, giving it NAND-class capacity. Intel also says that chip-for-chip, Xpoint is a thousand times more durable compared to 3D NAND. Of course, Xpoint is non-volatile like NAND, as well.

Xpoint is just one piece of the puzzle for Optane products, however. Intel has developed a custom controller to take advantage of the unique properties of the Xpoint material. For the DC P4800X, the controller talks over PCI Express using the NVMe protocol. Like an SSD controller, Optane products use multiple channels with multiple dies per channel to exploit parallelism for greater performance. However, Optane’s controller ASIC handles both data and commands exclusively in hardware to deliver the full performance of Xpoint media (at least the performance possible on the PCI Express 3.0 bus). The Optane controller can also spread a 4KB read operation over all of its dies, which Intel contrasts with the single-die read handling of NAND.

Thanks to Xpoint’s fine-grained access characteristics, Optane devices can also write in place on the underlying medium, unlike NAND’s read-modify-write cycle at the block level. In turn, Xpoint should match its higher performance with high endurance.

With a drive as fast as the P4800X, Intel believes that up to 30 drive writes per day is going to be adequate endurance for the types of applications it believes will benefit from Xpoint’s performance. With a rated maximum of 12.3 petabytes written, that means the 375GB DC P4800X drive could endure 30 DWPD for three years of constant use by our calculations. It’s worth noting that the projected endurance figures in the graphs above come from the 750GB P4800X, not the 375GB drive. Larger drives should enjoy a greater total petabytes written figure if Xpoint endurance scaling is similar to that of NAND. The best-case NAND endurance figure, on the other hand, comes from Intel’s DC P3700 800GB and its 14.6 PBW figure, good for 10 DWPD.

The 375GB P4800X uses a seven-channel controller with four dies of 3D Xpoint per channel for 28 dies total. Although Xpoint devices don’t require overprovisioning like NAND SSDs, the P4800X still reserves some extra space for ECC, firmware, and reallocation of bad regions of its total storage pool.

 

High speeds with light loads—and consistent performance nearly everywhere

Because of these properties, Optane products have exceptionally low latency for persistent storage devices, and it’s that characteristic—not sequential performance—that you’ll hear emphasized in Intel’s marketing materials for Optane. In fact, Intel won’t be providing sequential performance specifications for Optane SSDs in official materials, because it believes those figures will be misused by its competitors to misrepresent Optane SSD performance compared to NAND.

Intel justifies this stance by noting that NAND SSD makers (itself included) can only deliver their specified performance numbers by testing workloads with unrealistically high queue depths like 128. Outside of those synthetic situations, the company claims its internal testing shows that even heavy server workloads often don’t exceed QD16, and that will be especially true of Optane thanks to its symmetrical read and write performance. That symmetry means read requests won’t be held up behind higher-latency writes, as they would be with NAND controllers. In fact, Intel says that those who need large sequential numbers at high queue depths should just use NAND devices.

Instead, the Optane SSD DC P4800X really shines in low-queue-depth random workloads. Compared to even the DC P3700 NAND SSD, Intel says the P4800X offers eight times the performance of the NAND drive at QD1, and that lead still holds at about five times the random performance of NAND as queue depths increase. In fact, the P4800X saturates at about QD12, and Intel says that the only characteristic that changes with deeper queues on this drive is its service latency.

That’s not the end of the P4800X’s performance benefits, however. Not only does the Optane drive’s low-QD performance shine compared to NAND, the consistency of that performance could also be revolutionary compared to today’s SSDs. TR readers are already quite familiar with the importance of 99th-percentile performance thanks to our Inside the Second graphics-card benchmarking methods (themselves inspired by the use of that metric in server benchmarking), and Intel is deploying a similar measure to demonstrate the performance of the P4800X.

Intel has devised what it claims is a better method of demonstrating the unique benefits of Optane in those worst-case scenarios: the “write pressure” test, in which the latency for a read operation is measured while the drive is servicing a sustained quantity of write IOPS. The stairstep line in the rather complex chart above shows the random write IOPS rate in MB/s. The squiggly blue line is the average read response time in microseconds for the DC P3700, while the orange line near the x-axis demonstrates the average read response time for the DC P4800X.

Although it’s not shown in the write pressure graph, Intel says it saw similar average read response times for the P4800X all the way out to a remarkable 2GB/s of random write pressure. That consistency is a huge deal for Optane application performance, just as a low 99th-percentile frame time is for gaming. Even if one considers 99.999% response time for the DC P4800X on the logarithmic scale above, it’s still well below the average response time for the NAND DC P3700.

As just one example of the way the DC P4800X can boost application performance, Intel presented an example case of a two-server MySQL setup with a database too large to fit in RAM. Running Sysbench 0.5 with a 70-30 read-write split, and with the data stored on a 400GB Intel DC P3700, the rack could handle only 1395 transactions per second. With the 375GB DC P4800X, that number soars to 16480 transactions per second with similar 99th-percentile latency. That’s an impressive performance boost from mostly the same hardware.

 

Extending DRAM with Intel Memory Drive

Even though the DC P4800X is an NVMe device, Intel will begin providing a way to use it as a method of extending system memory while it develops Optane DIMMs for future systems.

This middleware, called Intel Memory Drive, serves as a hypervisor of sorts that will sit between the operating system and whatever complement of main memory and Optane devices are installed on the host system. Memory Drive will allow flexible provisioning of Optane-devices-as-storage and Optane-devices-as-memory to the system administrator. This technology requires no reprogramming of the operating system or applications. The Memory Drive and Optane suite will only work on Intel’s Xeon platforms.

Optane can extend the capacity of a system’s DRAM  thanks to its reliably low latency. That’s because paging to an Optane SSD doesn’t carry the uncertainty or nearly as much of the performance penalty of paging to NAND devices or a hard disk. To prove this point, Intel showed that GEMM, a type of workload commonly used in deep learning, can run (in a highly optimized form) at 2605 GFLOPS with 128GB of DRAM and 1.5TB of P4800X capacity installed, compared to 2322 GFLOPs with 768GB of DRAM alone installed. At $6080 for the Optane SSDs and roughly $2100 for a 128GB DDR4-2400 DIMM, the cost of crunching those large data sets could be more favorable with Optane, too. Achieving 768GB of DRAM with 128GB DDR4-2400 ECC DIMMs would require almost $13,000 of memory.

Even in less favorable workloads for Optane, extending DRAM with P4800X SSDs and Memory Drive could let companies trade some of DRAM’s performance for the ability to get huge data sets much closer to the CPU than DRAM capacities currently allow. Intel says that a two-socket Xeon system provisioned with DRAM alone could reach 3TB of caching space, while a combination of DRAM, Memory Drive and Optane P4800X SSDs would allow that system to pack in as much as 24TB of memory-class space. A four-socket system would double that figure to 48TB. Based on our calculations, those figures are only possible with some combination of 1.5 TB DC P4800X drives, but even if those drives scale linearly with the 375GB drive’s roughly $4-per-gigabyte cost, they could still economically boost the amount of memory-class cache close to the CPU in ways that DRAM alone can’t.

All told, the performance improvements in Optane lead Intel to call the DC P4800X the industry’s most responsive datacenter SSD, and we see no reason to contest the point. The 375GB DC P4800X will be available for $1520 starting today, and the drive should become more broadly available over the second half of this year. Intel will begin offering a 750GB version of the drive in the second quarter of this year, and a 1.5TB drive will follow in the second half of 2017. U.2 versions of these drives will also be available in the second quarter of this year for the 375GB version. U.2 versions of the larger-capacity drives will become available in the second half of the year.

Comments closed
    • DavidC1
    • 3 years ago

    If you have gotten in the initial batch of good SSDs starting with Intel’s X25-M, then its unlikely you’ll really need an upgrade.

    [url<]https://techreport.com/review/30813/samsung-960-pro-2tb-ssd-reviewed/5[/url<] Look at the game loading times. The X25-M G2 is there and its still competitive. The 960 Pro is often less than 0.5 seconds in the lead. That's like going from 2600K to a 3770K. But with 7 year difference. You thought only CPUs were advancing at a minimal pace. Theoretical specs put 960 Pro way ahead of the first gen good SSDs. In all metrics, latency, QD1, QD32, sequential, random, you name it. However, it does not pan out for real world. It tells me to get real benefits, we need to rip out everything and rebuild it from scratch. Probably software limitation is the culprit, because its still based on the mindset when HDDs were the king. We are talking something in line with revolution. Basic computer science teaches us loading into RAM from hard drives are everyday life. While it'll take time, it can start with less performance critical areas. What am I talking about? What people call end-game with 3D XPoint: Storage, and memory, all in one.

      • squngy
      • 3 years ago

      Who ever said this was for games?

      Optain is a super low latency drive, it’s great for making millions of small transactions. Games do the opposite, they load a few large files.
      You could make game load times as fast as on an SSD just by raiding enough HDDs, on the other hand, no matter how many drives you raid together you could never get to Optains low latency, not even if you used SSDs.

      What you’re saying is a bit like bringing a bazooka to a cooking contest and then complaining that it didn’t make the pie taste any better…

    • dikowexeyu
    • 3 years ago

    This is not a review

      • UberGerbil
      • 3 years ago

      You are correct. This is a technology brief / preview / backgrounder. TR publishes these regularly, whenever a major new architecture or tech comes along (here’s [url=https://techreport.com/review/30619/amd-unwraps-its-seventh-generation-desktop-apus-and-am4-platform<]one for Ryzen[/url<], for example, and [url=https://techreport.com/review/28189/inside-arm-cortex-a72-microarchitecture<]one for ARM's Cortex A72[/url<], and [url=https://techreport.com/review/31224/the-curtain-comes-up-on-amd-vega-architecture<]one for the Vega GPU[/url<]). These are not reviews. They set the stage for the reviews that come later, by providing a lot of supporting information that the review can refer to so that the review itself isn't burdened with a lot of background that most enthusiasts already know. And they're also necessary because the tech companies tend to disclose information this way -- first the background, then the review samples -- so to ignore the information until a review is possible is to fail at the job this website claims in its very title. Helpful hint: with storage products especially, you can tell which are reviews by [url=https://techreport.com/storage/<]the use of that word in the title of the piece[/url<].

        • dikowexeyu
        • 3 years ago

        The URL shouldn’t say “Review”
        [url<]https://techreport.com/review/31608/...[/url<]

          • UberGerbil
          • 3 years ago

          Maybe it shouldn’t. And maybe you shouldn’t be basing your expectations on the [i<]address[/i<] of the article but on its actual [i<]content[/i<]. [url=http://penisland.net/<]URLs[/url<] [url=http://speedofart.com/<]can[/url<] [url=http://lesbocages.com<]say[/url<] [url=https://www.dicksondata.com/<]all[/url<] [url=https://www.gotahoenorth.com/<]sorts[/url<] [url=http://www.whorepresents.com/<]of[/url<] [url=http://www.sbnation.com/2011/10/7/2474722/miami-marlins-ballpark-home-run-structure-what-is-that-gaaaaaaaaah<]things[/url<]. Who even pays attention to them, unless you're a semi-sentient google bot crawling the web?

    • Cannonaire
    • 3 years ago

    I dunno. What kind of tangible benefits am I going to see by installing [b<]SkiFree[/b<] on this?

      • chuckula
      • 3 years ago

      Clearly this product is overpriced for the SkiFree market.
      I would recommend buying two GTX-1080Ti cards for adequate performance instead.

        • EzioAs
        • 3 years ago

        So that the snow monster can get to me faster?! No thanks.

    • dodozoid
    • 3 years ago

    I wonder if the hybrid memory controler of AMD’s Vega would be compatible with this technology. I would fit its purpose much better than slugish NAND.

    • DarkUltra
    • 3 years ago

    Does this thing have capacitors to write out anything in the controller/buffer in case of power loss to prevent something like this:

    [url<]https://youtu.be/-Qddrz1o9AQ[/url<]

      • chuckula
      • 3 years ago

      I would pay that guy big bucks to run an IT operation.

      At my biggest competitor.

      • Waco
      • 3 years ago

      Yes. You won’t find anything even remotely close to an enterprise-class drive that doesn’t have that feature.

      • DavidC1
      • 3 years ago

      If it does it won’t need anything significant because the wear leveling would be far simpler. Plus it does not need DRAM back up to speed up writes because it does not need erase before re-write.

    • joselillo_25
    • 3 years ago

    It would be nice a new socket and motherboards with just a a CPU and this memory for the consumer market. No RAM and no HDD.

    • JosiahBradley
    • 3 years ago

    Edit: removed comment, only optane memory is locked in (still dislike that). Thanks Jeff for the explanation.

      • Jeff Kampman
      • 3 years ago

      Nothing about this product is “locked in” unless you want to use the Memory Drive tech. It acts like any other NVMe SSD.

        • JosiahBradley
        • 3 years ago

        Then why can’t I it into a z170 board or use a Skylake chip? This was stated by Intel you need certain chips to use it.

          • Jeff Kampman
          • 3 years ago

          Those limitations apply to Optane Memory, which is a separate product you’ll be hearing about soon.

            • chuckula
            • 3 years ago

            [spoiler<]OMG SPOILER ALERT KAMPMAN![/spoiler<]

    • slushpuppy007
    • 3 years ago

    At such impressive QD1 figures, this thing should load a game much faster than any NAND SSD.

      • ptsant
      • 3 years ago

      You are probably CPU limited by decompression and “de-serialization” of game assets when loading even with a SATA SSD. Going from a standard SATA SSD to the fastest NVM only shaves a few seconds in launch times. You would have to make your applications aware of the underlying memory architecture if you want to really optimize the load times. This quite hard, because traditional data structures are usually conceived for RAM or rotating disks.

        • Froz
        • 3 years ago

        I think games are developed with SSDs in mind more and more often. I was really surprised when I recently played Total War: Warhammer from HDD. Battle loading took at least 30 seconds, perhaps even 60 (googling now I see people with HDDs complain that for bigger battles it can take even few minutes). When I moved the installation to SSD, the time was shortened to 5, maybe 10 seconds. I don’t remember such huge differences in the games I played at the time I bought my first SSD a few years ago. I wish TR checked Total War in their SSD reviews tests, I wonder if there is a notable difference between slower and faster SSDs.

          • Klimax
          • 3 years ago

          Hm, how much RAM do you have?

            • Froz
            • 3 years ago

            Only 8 GB. Why do you ask, is that somehow connected?

            /edit: note that when I played the game on HDD, the system partition was still on SSD. So it’s not about swapfile.

            • UberGerbil
            • 3 years ago

            [quote<]/edit: note that when I played the game on HDD, the system partition was still on SSD. So it's not about swapfile.[/quote<]Except it still could be. Pages containing read-only data (including code) don't get put in the page file, because they can be read whenever needed from the source disk. If you're so low on total memory (RAM + Pagefile) that pages are getting evicted to make room for whatever is coming in, you could end up getting gated by HD read speed. I don't know what the working set of the game in question is, but that can be a bit moot during transient spikes like level loads anyway if the game isn't careful to keep core code/assets from getting discarded and then loaded back in from the HD afterwards. (And, to the earlier point, that may be the kind of thing developers and testers don't notice if they are SSD-only)

            • Klimax
            • 3 years ago

            Yes. The more RAM the more files stay in file cache and thus load on driver is reduced. That’s why I can keep using cheaper n TB drives (currenty4+6TB) for games without suffering long load times. Windows are very aggressive about caching files when there is load of spare RAM. (Aka what good RAM is when it sits unused…)
            Note: I got 16GB of RAM.

            • Froz
            • 3 years ago

            To be honest, I doubt having more RAM gives you the same results as just having the games on SSD. Sure, prefetching is nice and all, but it’s not going to load up the whole game to RAM, so at some point you will have to read from disk. For reference, TW: Warhammer takes 33 GB space on disk. So, yes, more RAM might help a little, but I seriously doubt it is going to make that huge difference. And would prefetching ever used it all up for just one application? I don’t know. Anyway, I would be happy to see someone test it. If one day I’ll get another 8 GB RAM or more, I might check it.

            • Klimax
            • 3 years ago

            I have 16GB. Only on initial start I see some severe loading on any game, after that it is barely noticeable. And if those files are actively accessed and there is no directive from code to not cache, they will be cached. And rest is covered by WD Black.

            So I can state that yes, there is fairly brutal difference. Even across exit/start.

      • Krogoth
      • 3 years ago

      You realize that almost every mainstream application out there is CPU-bound for load-times under a merger SATA SSD device?

      PCIe NVMe devices and beyond only begin to make sense if you are doing heavy-duty I/O workloads while gaming at the same time. The guy in the article only makes sense under an enterprise datacenter/HPC type of workload. It is suppose to be a low-cost alternative to ultra-high dense RDIMMs and LRDIMMs that are commonly found here without having to sacrifice latency and I/O throughput.

    • ronch
    • 3 years ago

    Hey look it’s the new GeFor… um..

    • Krogoth
    • 3 years ago

    This guy is a game-changer in the datacenter/HPC world.

    Impressive I/O throughput performance at its price-point.

      • morphine
      • 3 years ago

      There we go, folks! Krogoth is impressed, Optane is officialy for the win.

      • Beahmont
      • 3 years ago

      You know you’ll start to loose your reputation like this. Impressed by AMD and Intel in the same month? I’m seriously Krogothed at you right now Krogoth.

    • Wonders
    • 3 years ago

    Awesome rundown!

    • the
    • 3 years ago

    The improved latency and endurance seems to be there, at least on the marketing slides. So for those metrics, Optane does make sense so it does have a niche in the server market to occupy.

    The raw performance aspects have me a bit worried. Flat out I think Intel is saying that raw read/write bandwidth is not going to be leading class. I think chose the wrong controller path (seven channel?!?) versus a more traditional 8 channel + 1 channel for ECC approach that is well understood.

    The other downside is cost. There claim of $2100 for a 128 GB DIMM is a bit high as I’ve seen 3rd party (Crucial, Kingston, Samsung) memory floating around the $1700 to $2000 range. Granted, server vendor pricing (Dell, HPE, Lenovo) is likely higher so it isn’t an unjustified amount. At the lowend of $1700 per DIMM, one would need to speed ~$9600 for 768 GB of memory, not ~$13,000. This also puts Optane at barely half the $/GB as load-reduced ECC memory. I was expecting it to be even less. I also see that 1.1x speed increase with Optane purely due to the raw extra capacity vs. a lower capacity DRAM solution. The real apples-to-apples test would be to use similar capacities, see where the benchmarks result (I’d wager on DRAM) and let the difference in performance/$ dictate if using Optane is worth it.

    Benchmarking with differing capacities can be justified but only in specific contexts. The obvious is dealing with data sets beyond DRAM capacity (> 1.5 TB per socket). The other is a performance/$ analysis of a single socket system with Optane vs. a dual socket system with DRAM [b<][i<]and[/i<][/b<] software licensing. Here Optane could potentially offer twice the performance/$ due to the arcane manner of how enterprise software licensing works when a second socket is used. The flagship version of Optane was supposed to be the NVDIMM formate but [url=https://techreport.com/news/30869/intel-purley-server-platform-wont-use-3d-xpoint-memory<]it looks like that support has been removed[/url<] and this PCIe hardware/software solution put in its place. I also would have expected performance to be better than what Intel is announcing now.

      • ptsant
      • 3 years ago

      Thanks for your comment.

      It finally comes down, as always, to cost. Putting 512GB of Optane on your machine should cost a lot less than putting 512GB of RAM, otherwise the additional complexity is not worth it and RAM is much faster anyway.

      For example, 4×32=128GB ECC registered DDR4 can be found for $1200 in a local shop. In quad channel setups, this means 8×32=256GB for $2400 . The 375GB Octane is at $1572 MSRP.

      I’d rather get the RAM.

      PS. Hey, TR: how about running your benchmark on a … RAM drive?

        • the
        • 3 years ago

        There are a couple of inherent advantages of Optane over DRAM and NAND that give it a niche in the market.

        First is its non-volatile nature so everything should be able to survive a sudden shutdown. Essentially when power is restored, the system [i<]should[/i<] resume where it left off. This is the type of RAS feature that some environments will ask for even it is hurts performance. Considering that this type of system resume was not mentioned in the press release, I would suspect that this isn't supported. The latency and endurance advantages (on paper) over NAND shouldn't be under estimated either. Niches like high frequency trading will rapidly switch over this for bulk storage on that metric alone. The one thing that really does disappoint me is that the raw read/write figures should be exceeding NAND (or running into PCIe 4x limitations) and doing so while at lower queue depths. I hope that this is just a controller issue as Intel could easily add more channels for both capacity and performance. As I mentioned, the DIMM format has been delayed which could address some of these concerns. For the moment, Optane isn't a bad technology as it has its niche but more of a letdown from its earlier promises.

          • Bauxite
          • 3 years ago

          The shutdown thing can be nice, but server farms have UPS banks, dual power and generators for a reason. Granted a few years ago got to personally experience the cutover circuit for that hard fail 🙁

            • the
            • 3 years ago

            Yeah, until you’ve reached maximum power density for a data center. I’ve heard of some facilities drawing enough power to have rolling brown outs due to load.

            This would be some level of data protection on top of those other features. I’ve personally seen a data center go down during generator maintenance gone wrong so even if you have them, they can still fail.

      • Bauxite
      • 3 years ago

      Their pricing is the usual retail dell/hp w/o discount bullshit. (never ever pay list price for enterprise, if you are your procurement guy is probably a crook) Granted 64/128 modules are still load reduced which has a premium in price and a performance penalty as well.

      If you want to compare realistic builds, stick to capacity for a single socket because spreading the same workload around is usually a baaaad idea for performance. I personally put 256GB of 2400 DDR4 Registered ECC (8x32GB 2Rx4 type) in my workstation for $1500. (gone up a bit since last year…)

      Considering the only 375GB drive is launching at the same price and they haven’t announced something to beat the 6TB ram possible spread on a quad socket insane server, pretty fair to say there isn’t a lot of savings: roughly 33~50% cheaper than ram but quite a performance hit.

        • Waco
        • 3 years ago

        Lots cheaper than RAM if you’re looking at 128 GB LRDIMMs. They’re not cheap, even at discount pricing for quantity.

        That said, you also can’t put 30 or 40 of these in a server, so RAM still wins out there.

    • Rikki-Tikki-Tavi
    • 3 years ago

    I need this because of reasons.

      • Wirko
      • 3 years ago

      For the very same reasons, I need to know what the green and yellow blocks are made of.

        • derFunkenstein
        • 3 years ago

        Candy.

      • cygnus1
      • 3 years ago

      Get a couple for me too…

      • southrncomfortjm
      • 3 years ago

      Your reasons are inferior to my reasons.

    • Fonbu
    • 3 years ago

    Quote me if I am wrong. But wasnt the Optane products only compatible with Kaby Lake and above? Will this PCIe NVMe device work on any other systems?

    • derFunkenstein
    • 3 years ago

    Ah, now this product makes a ton of sense. I didn’t get the point of the initial tiny 32GB or 16GB offerings that Intel previously announced, especially since the performance wasn’t there (so using it as a cache didn’t even seem to have any benefit).

      • DavidC1
      • 3 years ago

      Performance is there for the market they are aiming at, which is an HDD accelerator.

      Besides, the lower sequential throughput does not change the fundamentals of the technology. The cache has the same ultra low latency as the datacenter version.

      The Kabylake only designation seems to be for the software bundle. Same with this Datacenter drive needing a Xeon for the Memory Drive.

        • derFunkenstein
        • 3 years ago

        Except that for those drives, the performance wasn’t really any better than cheaper, small SATA drives for the most part (large block-size reads excepted)

        [url<]https://techreport.com/news/30815/report-first-intel-3d-xpoint-products-coming-to-the-desktop[/url<]

          • DavidC1
          • 3 years ago

          I disagree. The cache device has one thing in common with this DC SSD.

          Optane SSD DC P4800X

          [url<]https://www.techpowerup.com/img/17-02-10/1b53ed9c3529.jpg[/url<] Sequential Read: 2.4GB/s Sequential Write: 2.0GB/s Random Read: 550K IOPS Random Write: 500K IOPS 550K x 4K = 2.2GB/s, 500K x 4K = 2GB/s The random numbers are same as sequential speeds. Optane Memory 16GB Sequential Read: 1.4GB/s Sequential Write: 300MB/s Random Read: 285K IOPS Random Write: 70K IOPS Optane Memory 16GB Sequential Read: 1.6GB/s Sequential Write: 500MB/s Random Read: 300K IOPS Random Write: 120K IOPS 16GB = 70K x 4K = 280MB/s, 32GB = 120K x 4K = 480MB/s, which is roughly equal to their sequential write throughput. The chips themselves may have a limit on bandwidth just like SSDs have ONFI standards for bandwidth and they use parallel access to increase it on the device. But the random write saturating sequential is key. The latency will be just as low as the DC version at 10us because its using the same technology. And latency is what beats the crap out of those SATA, even NVMe(because its not much of an improvement) NAND drives.

    • TwoEars
    • 3 years ago

    The 4k performance is unrivaled, can’t wait until this becomes mass-market and there are some decently priced consumer drives around.

    “Instead of working on pages with many kilobytes of data as NAND does, Intel says the structure of the Xpoint device allows for bit-level access.” That’s huge.

      • Waco
      • 3 years ago

      Pretty sure it’s byte-level, but yeah, no 128KB to 512KB page domains on these guys.

    • willmore
    • 3 years ago

    Nice writeup, Mr. Kampman.

    I question the acceptance that the Memory Drive software will gain in industry.

    While I appreciate how it removes the need for specific OS support (I’m looking at you Microsoft), I can’t help but think that using a sort of hypervisor to fake the reality of the situation is sub-optimal.

    On the plus side, I would expect that support for this type of memory partitioning will quickly become adopted in the open operating systems (Linux, the various BSDs) which are where it’s more likely to be used, so it might not end up mattering.

    Further on that point, I wonder how such memory will be addressesed. Will they treat it as a type of NUMA? That’s sort of confusing as NUMA gernerally is there to say “this memory is on my node and can be accessed quickly while *that* memory is further away”. Now, we’ll have to say “this memory may be on my node–and have plenty of BW–but will have a latency penalty as if it were on another node.” So, where we once though of latency and bandwidth both being dependent on distance from our node, it’s now become a bit more complex.

      • chuckula
      • 3 years ago

      [quote<]While I appreciate how it removes the need for specific OS support (I'm looking at you Microsoft), I can't help but think that using a sort of hypervisor to fake the reality of the situation is sub-optimal.[/quote<] I agree that in the long-term it's not the best approach, but for right now it serves as a stop-gap until mainstream operating systems can handle these memory hierarchies by themselves. For all I know, the Linux kernel + some tweaking could use it now without requiring this hypervisor since tuning for different types of NUMA configurations has already been done in the past.

        • willmore
        • 3 years ago

        Yeah, that’s what I was thinking. I haven’t been following that segement of kernel development, so for all I know, it’s already in there. But, yes, I expect Linux to be the first OS to the bar with full support for this type of use.

          • Klimax
          • 3 years ago

          There’s not much missing in NT Kernel either. Mostly UEFI firmware needs to be updated and few config options for memory manager to know which pages are not to be zeroed out. Everything else is already impale and I would say for more then decade. (if not decades)

      • DavidC1
      • 3 years ago

      The server usage cases are numerous. There will of course be cases as you mentioned where its not optimal. But many that are.

      What says they aren’t in use already?

      • davidbowser
      • 3 years ago

      My problem is that I am having trouble with the application level of this for the same reasons you mentioned. If I am designing something new today that requires some big dataset, I am looking at horizontally scale-able in-memory systems first.

      If I am updating an existing application with an existing SQL solution, and this looks good, I have all the OS level limitations that come with this. So I guess I am comparing this to shared storage multi-tier data caching systems, and wondering where the advantage of this is.

      What am I missing?

    • UberGerbil
    • 3 years ago

    I have to say when assembling a (i3-7100) NUC recently I was kind of amused to see it declared to be “Optane-ready” in the specs. Apparently Intel expects even a lowly i3 to be sporting these things eventually. It’s always interesting to see which techs Intel wants to push so hard that they put support into every platform, vs the ones they prefer to hold back to create market segmentation.

    [sub<]Edit: specified the NUC CPU[/sub<]

      • Ninjitsu
      • 3 years ago

      OTOH that means they intend to get the prices down on these things.

      • DavidC1
      • 3 years ago

      So, the bigger new NUCs support regular hard drives. Optane-ready points to the caching devices, the 16-64GB versions leaked previously.

    • Chrispy_
    • 3 years ago

    Woooo, it’s here.

    Now we just need to wait for a year until the big server OEMs offer it at anything [i<]remotely approaching[/i<] those prices.

    • brucethemoose
    • 3 years ago

    So only 3x the durability of binned MLC?

    If this is the future of the “unified memory” PC, I hope they can get those endurance figures up a bit…

      • Airmantharp
      • 3 years ago

      …especially given that the next major step is to shove them in DIMM slots.

      I can imagine that these drives may be wearing themselves out quicker as more unique workloads are found, but I can also imagine that if they make a commercial difference burning through them quick may still be cost effective.

      • Beahmont
      • 3 years ago

      3 times the endurance of special binned MLC as it’s normal base endurance.

      I’m not even sure with the manufacturing process used you [i<][b<]can[/b<][/i<] bin for better durability given we know next to nothing about what this stuff even is, let alone how it's manufactured or binned now. That said, we're already talking about time and endurance ratings that would have these things running for almost a decade, which is likely longer than most of the systems it will be put into in practice.

    • chuckula
    • 3 years ago

    FYI, you’ll note that Intel refers to the “DC P3700” as a reference point in this article. [Edit: Incidentally, if the photos in the article are true, the new P4800X in this article uses a PCIe 4X connector similar to many other PCIe SSDs]

    TR has reviewed the DC P3700 here: [url<]https://techreport.com/review/28032/a-fresh-look-at-storage-performance-with-pcie-ssds[/url<] And The DC P3700 was one of the drives used in the more recent review of the Samsung 960 Pro that you can see here: [url<]https://techreport.com/review_full/30813/samsung-960-pro-2tb-ssd-reviewed[/url<]

    • chuckula
    • 3 years ago

    [quote<]To prove this point, Intel showed that GEMM, a type of workload commonly used in deep learning, can run (in a highly optimized form) at 2605 GFLOPS with 128GB of DRAM and 1.5TB of P4800X capacity installed, compared to 2322 GFLOPs with 768GB of DRAM alone installed. At $6080 for the Optane SSDs and roughly $2100 for a 128GB DDR4-2400 DIMM, the cost of crunching those large data sets could be more favorable with Optane, too. Achieving 768GB of DRAM with 128GB DDR4-2400 ECC DIMMs would require almost $13,000 of memory.[/quote<] Thanks for doing a little number crunching to emphasize the point. People see $1500 for a 375GB drive and they think it's a ripoff. That's not the point of these drives, and even in at this rather primitive stage (the *real* long-term goal is to have Optane DIMMs that aren't PCIe constrained) there's already some strong potential to drastically reduce the price of a large-memory server while maintaining decent performance.

    • Waco
    • 3 years ago

    I don’t mind these longer news-type articles at all. Keep up the good work!

Pin It on Pinterest

Share This