Intel pumps the Optane SSD DC P4800X to 750 GB

Optane SSDs have left a good impression here at TR. Some folks sneer at the performance claims Intel makes for them, and it's true that the sequential operation figures aren't shocking. The Optane drives are unmatched in terms of responsiveness, though—much the same way early SSDs weren't a lot faster than HDDs in sequential terms, yet blew them away in random performance. The fastest solid-state drive on the planet right now is Intel's Optane SSD DC P4800X, and the boys in blue just announced that you'll soon be able to buy a double-capacity version storing 750 GB.

The SSD DC P4800X normally comes in what is sometimes called an HHHL form factor, referring to a half-height half-length PCI Express card. However, the existing 375 GB size isn't a lot of capacity for the physical space required, so doubling the storage should make sticking a few P4800Xes in your servers a less-painful proposition. If you're really strapped for space, Intel is also about to start offering the P4800X in a 2.5" form factor with a U.2 connection. U.2 is basically "M.2 with a cable," so the performance of these drives should be identical.

Intel didn't say how much the double-stuffed Optane SSD or the U.2 variants would be when they hit the market soon. The existing 375GB SSD DC P4800X goes for $1730 on Amazon, so the new drives will probably be well into "if you have to ask" territory.

Comments closed
    • Mr Bill
    • 2 years ago

    I wonder if configured it to a much wider memory bus, this could proxy for HBM in video cards.

    • DavidC1
    • 2 years ago

    PC Perspective has also shown that RAM Drives are only 3.5x faster in terms of latency(thus IOPS).

    That means a RAM Drive is at ~3us or so and the Optane drives at 10us. RAM without “RAM Drive” is an additional 50x faster.

    That puts into perspective how *everything* needs to be changed to have an NVRAM device that’s in the DRAM like latency. Current systems won’t cut it.

      • Klimax
      • 2 years ago

      No. That overhead will have to stay. How else do you keep track of files, permissions (…) on storage? Magic?

      And no unified storage/workspace case either. Unless you like corrupted data and/or extra overhead…

        • DavidC1
        • 2 years ago

        No, it doesn’t.

        It’s up to the engineers and their management to figure out how to abstract the hard parts as much as possible from the users.

        For over 40 years it has been about loading to RAM because storage was too slow. Running it off storage would have been unrealistic with the performance loss.

        EVERYTHING has to change. Even the mindset of programmers. 15-20 years from now there will be computers that should have unified RAM/Storage. Those systems would turn the “load from Storage to RAM” concept on its head. That’s how you bring the next revolution in computers. No loading, no sleep modes, instant on. Man, even an Average Joe would see the point of replacing current systems with such ones.

        If you think current file systems and the way OS handles things make it “impossible”, then those concepts have to go too.

          • Klimax
          • 2 years ago

          Completely missed all the things. Just because underlying HW changes doesn’t mean original principles and requirements disappear. You focus only small part of entire system, missing everything else and that everything else is critical part of equation. (Otherwise one is in deep trouble)

          You always need structure to describe location of files and their attributes, you always need security descriptors, you always need separate workspace from storage and so on.

          Skipping any means lost and corrupted data and/or giving them over to adversary.

          I wasn’t surprised by your ignorant comment. It is sadly too common. Doesn’t mean it is any less wrong.

          ETA: Reminder: If there is no separation of storage and workspace, any error in code or system can and will lead to data loss. And there is no way around that problem.

          ETA: Anyway, your assertions about filesystems and their designs are at least in case of Windows quite wrong. (NTFS nor ReFS have any assumptions about storage built-in, storage part of equation is handled by different components)

    • chuckula
    • 2 years ago

    First of all: No, it won’t make your games load faster.

    The performance levels of these parts are a little scary when you think about a few points. Number one is that at least in this generation of products Intel has put in exactly zero RAM to buffer the I/O requests like you see in basically every high-end and most mainstream flash drives. That basically means that in most of the areas where Optane appears to not really be better than “flash” you probably are actually comparing Optane to the performance of the DRAM cache in the NVME drives.

    Second, PC Perspective [url=https://www.pcper.com/reviews/Storage/Intel-Optane-SSD-DC-P4800X-750GB-Review-Flesh/Bridging-Gap-and-Polling-vs-IRQ<]did some interesting testing that broke down the actual sources of delay in the service time for I/O operations in Windows[/url<] using both DMA mode (the "modern" standard way to service I/O via interrupts) and the older polling I/O mode that was around before DMA became widespread. Interestingly enough, the old polling mode actually showed benefits over DMA because the Optane drives are so fast that the polling overhead (normally a killer for regular drives) drops significantly while the fixed servicing delay for DMA interrupts actually becomes a noticeable factor in performance overhead. Basically, in at least some situations these drives are bottlenecked by the inherent I/O architecture of the PC with the interrupt request service overhead and not by the Optane chips themselves. Back when DMA first became widespread, we were using spinning hard drives as the metric for what "fast" storage actually entailed, and a few microseconds to service an interrupt request was nothing more than a rounding error. Even with high-speed SSDs it's still not that big of a deal. Now with Optane, we have exposed a new performance bottleneck that was never an issue in the past.

      • tsk
      • 2 years ago

      How do you remove that bottleneck?
      Build something other than DMA or PCIe?

        • chuckula
        • 2 years ago

        It probably comes down to removing the standard PCIe bottleneck or figuring out how to implement DMA with lower overheads. Part of the DMA issue is hardware and part of it is software since the OS in question (Windows) has to do a context switch to service the DMA interrupt. Making that operation faster could mitigate the issue.

          • tsk
          • 2 years ago

          Can EMIB be used to connect storage directly to the CPU? I’m not up to date in the storage world, but I’ve heard rumors of Intel looking to replace PCIe.

          • RAGEPRO
          • 2 years ago

          I imagine [url=https://techreport.com/news/30790/gen-z-consortium-prepares-for-the-future-of-storage-and-memory<]Gen-Z[/url<] is at least in part targeted at resolving this problem.

        • jts888
        • 2 years ago

        The topic comes up repeatedly over the years, but the only real solution is probably something along the lines of adding explicit messaging to ISAs like x86_64. As it stands, peripheral devices can push and pull blocks of memory to and from each other directly, while host CPUs just dump stuff to system memory and fire interrupts (a.k.a. “doorbells”) telling the peripheral when to start DMA’ing.

        Latencies could be cut very substantially if there was a uniform asynchronous communication model extended to CPUs, but this would involve at the very least bolting hw queue management units onto cache subsystems that would never (or at least not soon) really be beneficial outside enterprise storage and networking. And it’s something that you really need everyone on board with doing, and Intel has been making a lot of peripheral makers uncomfortable with their proprietary shifts in storage, networking (OmniPath), etc.

      • freebird
      • 2 years ago

      The PRICE is what scares me…

        • psuedonymous
        • 2 years ago

        Launch prices for NAND-based SSDs were insane. Launch prices for [i<]consumer[/i<] NAND SSDs were also crazy. Compared to those, the consumer Optane drives are hilariously cheap. Take the X-25M for example, the first not-awful consumer SSD (and already using MLC NAND to do it). $600 for 80GB. $600 today buys you 480GB of 3D XPoint, with much less time between DC-only drive availability and consumer drive availability.

      • Waco
      • 2 years ago

      The spinlocks versus interrupt argument is pretty well solved for things that aren’t IO. I’m glad to see people actually putting some thought into storage stack latency!

        • chuckula
        • 2 years ago

        I don’t think that PC Perspective was trying to say that going back to polling is the solution to the problem per-se since obviously that’s not an efficient solution outside of their specific tests. It was more of an illustration as to where bottlenecks can occur in the standard DMA setup.

          • Waco
          • 2 years ago

          For small block low latency stuff, it makes sense. For bulk transfers DMA is definitely the way to go.

            • chuckula
            • 2 years ago

            It reminds me of work I’ve done in packet capture analysis using a modified Linux network stack that puts packets in a ring buffer for high-speed capture since the default interrupt-based stack starts to keel over at higher bandwidths. [url<]http://prod.sandia.gov/techlib/access-control.cgi/2015/159378r.pdf[/url<]

Pin It on Pinterest

Share This