Slashdot Mirror

← Back to Stories (view on slashdot.org)

Fujitsu Picks 64-Bit ARM For Post-K Supercomputer (theregister.co.uk)

Posted by msmash on from the new-hardware,-new-hopes dept.

An anonymous reader writes: At the International Supercomputing Conference 2016 in Frankfurt, Germany, Fujitsu revealed its Post-K machine will run on ARMv8 architecture. The Post-K machine is supposed to have 100 times more application performance than the K Supercomputer -- which would make it a 1,000 PFLOPS beast -- and is due to go live in 2020. The K machine is the fifth fastest known super in the world, it crunches 10.5 PFLOPS, needs 12MW of power, and is built out of 705,000 Sparc64 VIIIfx cores.InfoWorld has more details.

20 of 30 comments (clear)

  1. Re:Frits! by __aaclcg7560 · · Score: 1

    Why would you want a Beowulf cluster of Cyrix processors? :P

    https://en.wikipedia.org/wiki/Cyrix_6x86

  2. Re:OT: Slashdot's MOTD Is Wrong by Hognoxious · · Score: 1

    Nah. Yogi Berra said "yabadabadooooo!"

    --
    Confucius say, "Find worm in apple - bad. Find half a worm - worse."
  3. nVidia? by LWATCDR · · Score: 1

    I wonder if nVidia might not get into this space as well. They make ARM cpus and are leaders in the GPU compute space. I tightly coupled ARM/Pascal system could be very interesting. The key would be the interconnects.

    --
    See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
    1. Re:nVidia? by Blaskowicz · · Score: 1

      They concentrated on interconnects that are of use inside a node and are baked in the motherboard pretty much. So that's something like Intel QPI, but not a supercomputer interconnect. For the near future you'll have to ask IBM what they use on POWER9 systems.

  4. Interesting choice by serviscope_minor · · Score: 3, Interesting

    Sounds like the core is a completely custom ARM job which is more or less the upgraded SPARC core back end with an ARM instruction decoder rather than a SPARC one. And with lots of goodies integrated, specifically a very fast interconnect (TOFU) and a large, fast, wide floating vector point unit.

    If they stick enough FPU grunt into the VFP unit, they won't need GPUs or coprocessors in addition.

    --
    SJW n. One who posts facts.
    1. Re:Interesting choice by Anonymous Coward · · Score: 1

      It's probably hard to justify the cost of continued SPARC development since they are the only vendor and serious developer of SPARC systems.

      Well there was Sun.. Who was using Fujitsu designed systems for all of their "big" stuff near the end. Sure, Oracle bought them.. But who would want to be Oracle's business partner? Oracle is famously a bunch of freeloading cheepscates. They don't even care to partner with Redhat for their enterprise Linux products. They just repackage Centos and brand it Oracle.

      I imagine Fujitsu didn't care to foot the bill for Oracle system development, and didn't see much of a future for SPARC when they can reap the benefits of the already quite rich ARM development community.

  5. Someone help me to understand... by SkyLeach · · Score: 1

    The cost of these supercomputers is astronomical, so why the petaflops?
    As I understand it the benefit of a supercomputer is BUS speed (getting data between the CPU caches quickly) for massively parallel computing tasks. What I don't understand is why there are so few hybrids using parallel GPU processing (2x512x8/16/32GB) to achieve the same tasks, even for weather applications. It seems to me that processing blocks could be exchanged over fiber sufficiently for real time applications, but I may not fully understand the problem?

    As I understand it the supercomputers have separate dedicated channels for types of memory exchanges for large matrices that get updated by all the processors, but block updates aren't read anywhere near as quickly as the chunks being worked on. It seems that the delay wouldn't make much difference overall, but I don't claim to be an expert on it.

    --
    My $0.02 will always be worth more than your â0.02, so :-p
    1. Re:Someone help me to understand... by serviscope_minor · · Score: 3, Insightful

      As I understand it the benefit of a supercomputer is BUS speed (getting data between the CPU caches quickly) for massively parallel computing tasks.

      Well, more latency than raw speed. And that's a question of the interconnect. The old SPARC chips had (and the ARM chips will have) the interconnect in the CPU die for minimum latency.

      What I don't understand is why there are so few hybrids using parallel GPU processing (2x512x8/16/32GB) to achieve the same tasks,

      A GPU basically couples a lot of FPU grunt with a really simple processor. The K-computer had custom CPUs which had somewhat simple (compared to Xeons) processors with huge vector floating point units attached, which achieves much the same effect. You don't get quite as much FPU grunt for the money, but you get something easier to program and since it's part of the CPU, it also connects straight to the embedded interconnect.

      --
      SJW n. One who posts facts.
    2. Re:Someone help me to understand... by SkyLeach · · Score: 1

      But I thought that the shader pipelines were basically low-latency vector processing chains.

      --
      My $0.02 will always be worth more than your â0.02, so :-p
  6. Who will make the chips? by TheRaven64 · · Score: 2

    The big question is whether this means that Fujitsu is going to start producing ARMv8 chips. They've got a lot of in-house expertise in SPARCv9 processor design, much of which would be directly applicable to building an ARMv8 pipeline (no register window weirdness, but there are a lot of similarities). The main difference would be that they'd be sharing the ecosystem (OS, compiler, and so on) costs with companies like ARM, Apple, Google, and so on rather than with Oracle, which sounds like a very good deal.

    --
    I am TheRaven on Soylent News
    1. Re:Who will make the chips? by serviscope_minor · · Score: 2

      The big question is whether this means that Fujitsu is going to start producing ARMv8 chips.

      Fujitsu already produce ARM chips, so my guess would be that they will make these ones too.

      --
      SJW n. One who posts facts.
    2. Re:Who will make the chips? by TheRaven64 · · Score: 1

      That makes sense. If they're already an ARM licensee, then it should be a fairly simple transition and there's a big benefit in depending on a compiler that a load of other people also depend on, especially if they're people with deep pockets.

      --
      I am TheRaven on Soylent News
    3. Re:Who will make the chips? by PingPongBoy · · Score: 1

      I wouldn't mind one of those ARMv8 in my laptop. I just have a 4 cylinder, and oil is so cheap right now.

      --
      Know your pads. One time pad: good for cryptography. Two timing pad: where to take your mistress.
  7. exaflop on less than a gigawatt by peter303 · · Score: 3, Informative

    Scaling up the K computer to an exaflop at one gigaflop per watt would require a gigawatt of power. The Green500 lists says the most efficient super runs at 7 gigaflops per watt in 2015. There are plans to do a lot betterbthan this.

  8. There is no leader by info6568 · · Score: 1

    When you read that the fastest machine is the Sunway based on chinese made CPUs (because of US restrictions), and that Fujitsu (Japanese) is working on so powerful machine based on ARM (England) designed architecture Fujitsu improved, it feels as an indicator that US is not leading the technology anymore.

    I will put this in different words: no one country can lead the future of any other country just because of past achievements. Now, there are several options and all them are valid.

    Also, following the Russia attempt to have backdoors on the encryption algorithms for messaging, it is clear that nobody is accepting any restriction from anybody else and that technology alone can't be a deterrent.

    As a side note: Could be possible that, because of political reasons, we are arriving to the end of technology advancement?

    1. Re:There is no leader by malditaenvidia · · Score: 1

      Could be possible that, because of political reasons, we are arriving to the end of technology advancement?

      How so? All this competition will only further engineering and development.

    2. Re:There is no leader by ceoyoyo · · Score: 1

      Apple, Intel, Marvell and others had quite a bit to do with modern ARM, as well as UK-based Acorn. On the other side, if you're using an Intel processor it's probably based on an architecture designed by an Israeli team.

      Most big projects like designing processors or building supercomputers are really international. You might be right though, the US might have chased away some of their share through export restrictions.

    3. Re:There is no leader by ogdenk · · Score: 1

      And don't forget DEC. They made the StrongARM SA110 which was the first high performance ARM. Hit 233MHz back when PC's were usually running at a similar clock speed. It's what made the later Apple Newton 2x00 series not suck.

  9. Re:OT: Slashdot's MOTD Is Wrong by SScorpio · · Score: 1

    That was Fred Flintstone. Yogi said "Hey Boo Boo".

  10. Re:Frits! by malditaenvidia · · Score: 1

    Because he is an arsonist, okay? He can't help himself, stop judging the poor man.