Stories
Slash Boxes
Comments

SoylentNews is people

posted by FatPhil on Tuesday September 28, @01:24AM   Printer-friendly [Skip to comment(s)]
from the chip-crisis?-what-chip-crisis? dept.

Blazing fast PCIe 5.0 SSD prototype hits sequential read speeds of 14,000 MB/s:

Advancements in the storage segment are the unsung heroes in today's world of computing. While many users tend to focus on the speed of their CPU, GPU or even the higher refresh rate of their displays, the increasingly quick solid state drives are in part responsible for the performance improvements of Sony's and Microsoft's next-gen consoles. But while the PlayStation 5 and Xbox Series X/S rely on PCIe 4.0 SSDs, a Japanese memory manufacturer is already finalizing the development of its blazing fast PCIe 5.0 storage solutions.

In a recent presentation, Kioxia has now revealed how quick PCIe 5.0 SSDs can truly be. While the throughput of the PCIe 5.0 interface at 32GB/s per lane is exactly twice as high compared to PCIe 4.0, the company's first prototype has apparently reached sequential read speeds as high as 14,000MB/s. That is also twice as fast as Kioxia's currently top of the line PCIe 4.0 drive.

Even though these read speeds certainly seem impressive, the write speeds of Kioxia's PCIe 5.0 SSD are similarly spectacular. The official benchmark says the drive can reach sequential write speeds of 7,000MB/s, which is a 67% improvement to the predecessor. Overall, these speeds seem to be absolute overkill for most use cases, which is why these drives are intended for use in a professional server setting. Nevertheless, the rapid advancements in storage speeds certainly deserve more attention than the often incremental upgrades in the CPU and GPU sector.

Of course, there's more to storage than speed, there's reliability, for example. Would any gamers want to prove the "professional server setting" assumption wrong?


Original Submission

Display Options Threshold/Breakthrough Reply to Article Mark All as Read Mark All as Unread
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
(1)
  • (Score: -1, Troll) by Anonymous Coward on Tuesday September 28, @01:30AM

    by Anonymous Coward on Tuesday September 28, @01:30AM (#1182072)

    blow that snow down below ahhh feels good!

  • (Score: 3, Interesting) by Anonymous Coward on Tuesday September 28, @01:35AM (5 children)

    by Anonymous Coward on Tuesday September 28, @01:35AM (#1182075)

    The question to ask is, then, what write speed the drive can SUSTAIN once its fast buffer is used up?

    The peak value in a contrived setting means nothing at all. A drive in a real-world system must handle real-world loads, and be fast doing THAT.

    • (Score: 2) by krishnoid on Tuesday September 28, @03:49AM (2 children)

      by krishnoid (1156) on Tuesday September 28, @03:49AM (#1182096)

      I'm more curious about speed comparisons to the other components on a typical PC motherboard. I mean if the SSD blows past memory and/or CPU speeds, wouldn't it mean current PC architecture [anandtech.com] could start to bottleneck performance?

      • (Score: 2, Informative) by fustakrakich on Tuesday September 28, @04:01AM

        by fustakrakich (6150) on Tuesday September 28, @04:01AM (#1182097) Journal

        Well, now you might need a bigger heat sink for the SSD than the CPU..

        --
        Ok, we paid the ransom. Do I get my dog back? REDЯUM
      • (Score: 2) by maxwell demon on Wednesday September 29, @09:33AM

        by maxwell demon (1608) on Wednesday September 29, @09:33AM (#1182681) Journal

        I guess it could mean that swapping gets far more attractive, as now your computer won't slow down to a crawl due to disk I/O.

        Note that the slowest DDR4 speed [crucial.com] is 19200 MB/s, which is in the same ballpark but still faster than the quoted SSD read speed. and about 2.5 times the quoted SSD write speed. The SSD would outperform DDR3 memory, but I guess if you invest the money to buy this disk, you'll not settle for DDR3 memory.

        --
        The Tao of math: The numbers you can count are not the real numbers.
    • (Score: 2) by Common Joe on Tuesday September 28, @07:27PM

      by Common Joe (33) <reversethis-{moc ... 1010.eoj.nommoc}> on Tuesday September 28, @07:27PM (#1182397) Journal

      You get a point from me, AC, because that was my first thought. We just had an article a few weeks ago (too lazy to look it up) where the companies are playing around with speed numbers and under delivering in the real world.

    • (Score: 0) by Anonymous Coward on Thursday September 30, @06:05PM

      by Anonymous Coward on Thursday September 30, @06:05PM (#1183142)
      Well if you have 128GB RAM it'll still take a while to hibernate and resume from hibernation.

      What? The rest of you don't have multiple browsers with zillions of tabs? ;)
  • (Score: -1, Troll) by Anonymous Coward on Tuesday September 28, @01:39AM

    by Anonymous Coward on Tuesday September 28, @01:39AM (#1182077)

    Gur onpba vf vafvqr gur pnxr

    Lbh whfg jnfgrq gvzr gung pbhyq'ir orra fcrag qbvat abguvat

    tG6C H2?E65 E@ A:DD 2== @G6C J@FCD6=7n (6== ?H D E96 E:>6P }~(P

  • (Score: -1, Troll) by Anonymous Coward on Tuesday September 28, @01:43AM

    by Anonymous Coward on Tuesday September 28, @01:43AM (#1182078)

    Just lube it up and you've got it! Your own private sneakernet storage system! Now find another person's ass to hide it in, like "passing the torch" until you get to your final source!

  • (Score: 2) by bzipitidoo on Tuesday September 28, @03:33AM (4 children)

    by bzipitidoo (4388) Subscriber Badge on Tuesday September 28, @03:33AM (#1182095) Journal

    Thought SSDs made read order almost entirely irrelevant to read speed. It's why they are so much faster than HDDs--- no seek time.

    • (Score: 3, Informative) by krishnoid on Tuesday September 28, @04:48AM (1 child)

      by krishnoid (1156) on Tuesday September 28, @04:48AM (#1182102)

      They do, but as a result, tagged command queueing [wikipedia.org] somehow still stayed valuable:

      • With hard disk drives that are really slow, having a queue of requests means the block requests could be satisfied in one or two passes by reordering them to follow the path of a single drive head sweep while the platter was rotating
      • With solid state drives that are really fast, having a queue of requests means multiple block requests can be satisfied immediately if the memory/controller is fast enough, since it's all random access anyway

      I think there's some fundamental wisdom about the "queueing" concept here, but I don't know excatly how you'd word it.

      • (Score: 1, Interesting) by Anonymous Coward on Wednesday September 29, @07:16AM

        by Anonymous Coward on Wednesday September 29, @07:16AM (#1182652)

        Almost all disks now use NCQ not TCQ. NCQ is important because SSDs are so fast that instead of the host waiting on the drive the drive often has to wait for the host. That leads directly into the major improvement that SSDs offer by using NVMe is that you can now have multiple queues that are much longer than with HDDs. This means that they don't have to sit and wait for the host as much and also allows for the SSD to run almost all actions concurrently and make more operations parallel because the requests made can match the internal architecture of both the SSD and the host system it is in better.

    • (Score: 2) by fraxinus-tree on Wednesday September 29, @09:44AM (1 child)

      by fraxinus-tree (5590) on Wednesday September 29, @09:44AM (#1182684)

      Sorry to say it, but the popular substandard wear-leveling layers made the sequental/random distinction relevant again. And what is worse, the write order matters as well (when reading). It is high time to trash the block-device concept for good and expose the nand directly to a reasonable filesystem driver. What we now do is patching both the wrong tools for the task (the block layer and the block-related fs layer) with things like "trim" and "nvme".

      • (Score: 1, Interesting) by Anonymous Coward on Thursday September 30, @10:08PM

        by Anonymous Coward on Thursday September 30, @10:08PM (#1183211)

        Sadly it is a bit more complicated than that. Back when SSDs came out, they did that directly without the FTL. But that caused problems because drives would lie, different types of drives require different treatment, and many admins would treat them like HDDs anyway. They overcorrected by making overcomplex FTLs but ended up hiding too much too. Now I think they are going to find a better middle ground, through the ZNS and its kernel-side zone file systems becoming more common as is the understanding that the device and the kernel can cooperate to their mutual benefit without stepping on each-other's feet.

  • (Score: 4, Interesting) by Rich on Tuesday September 28, @10:39AM (2 children)

    by Rich (945) on Tuesday September 28, @10:39AM (#1182152) Journal

    Had to look up what signaling is used on the wires. Wiki says 32GBaud (with 130/128 encoding). This bitrate means that the interface is 4-lane (out of possible 16). 32GHz is really a lot. Less than a centimeter of light wavelength. You probably can forget much of your electronics education for that and have to use special methods for guided radio signals. (Although they still use differential signaling).

    The drive side controller can have a single DDR4-4800 for those speeds, but for sustained transfers they'd have to smarten up the path to the flash cells. It wouldn't be black magic to have lots of parallel controllers for that, but someone has got to do it in the first place. Kioxia is Japanese (ex Toshiba Storage) and these like to have only incremental improvements to keep their long-term market going. With the basic technical blocks in place, a x16 bus and a wider backend could do four times as much, if datacenter throughput was really the main aim. Also, the presentation is in Chinese, which gives you an idea about the target audience. So "professional server setting" just means "we'll make it expensive until we can't anymore".

    • (Score: 2) by FatPhil on Tuesday September 28, @05:25PM (1 child)

      by FatPhil (863) <pc-soylentNO@SPAMasdf.fi> on Tuesday September 28, @05:25PM (#1182309) Homepage
      Baud != Hz
      --
      I know I'm God, because every time I pray to him, I find I'm talking to myself.
      • (Score: 3, Insightful) by Rich on Tuesday September 28, @10:59PM

        by Rich (945) on Tuesday September 28, @10:59PM (#1182472) Journal

        Baud = Symbols/Second. Up to PCIe 5 they use a scrambled NRZ variant, which packs a bit per cycle. Its 32 GBaud effectively equal 32 GHz in required line frequency. Only for PCIe 6, they changed to PAM signaling, again at 32 GBaud, but with four levels to pack two bits for 64 Gbps.

(1)