Catch up on stories from the past week (and beyond) at the Slashdot story archive

 



Forgot your password?
typodupeerror
×
Japan Medicine

Japan's Fugaku Supercomputer Goes Fully Live To Aid COVID-19 Research (japantimes.co.jp) 19

Japan's Fugaku supercomputer, the world's fastest in terms of computing speed, went into full operation this week, earlier than initially scheduled, in the hope that it can be used for research related to the novel coronavirus. From a report: The supercomputer, named after an alternative word for Mount Fuji, became partially operational in April last year to visualize how droplets that could carry the virus spread from the mouth and to help explore possible treatments for COVID-19. "I hope Fugaku will be cherished by the people as it can do what its predecessor K couldn't, including artificial intelligence (applications) and big data analytics," said Hiroshi Matsumoto, president of the Riken research institute that developed the machine, in a ceremony held at the Riken Center for Computational Science in Kobe, where it is installed. Fugaku, which can perform over 442 quadrillion computations per second, was originally scheduled to start operating fully in the fiscal year from April. It will eventually be used in fields such as climate and artificial intelligence applications, and will be used in more than 100 projects, according to state-sponsored Riken.
This discussion has been archived. No new comments can be posted.

Japan's Fugaku Supercomputer Goes Fully Live To Aid COVID-19 Research

Comments Filter:
  • by FudRucker ( 866063 ) on Friday March 12, 2021 @09:53AM (#61151218)
    yes., it runs Linux (Redhat Ent 8)
    https://www.fujitsu.com/global... [fujitsu.com]
    • by NFN_NLN ( 633283 )

      Thanks for the information. I was really worried IBM wasn't being compensated through support payments.

    • This says that it has "assistant cores". I can kind of guess what they might do, can anyone say a few words ?

  • Not mentioned in this summary is the fact that this supercomputer is built entirely using several million ARM chips as opposed to GPUs or x86. The computer is capable solving general problems rather than specialized vector-oriented problems. It's a demonstration that ARM provides a significant cost savings in up-front hardware and ongoing overhead (electricity) vs. the previous supercomputing architectures built with GPU and x86 hardware.
    • Re:built with ARMs (Score:5, Informative)

      by serviscope_minor ( 664417 ) on Friday March 12, 2021 @10:16AM (#61151282) Journal

      Not mentioned in this summary is the fact that this supercomputer is built entirely using several million ARM chips as opposed to GPUs or x86.

      No not really. Not mentioned in the summary is it's built using several million ARM cores as opposed to several million SPARC ones. Fujitsu has always used their own custom silicon and they switched from SPARC to ARM on this one. x86 and GPUs have never been in the running for this machine.

      The computer is capable solving general problems rather than specialized vector-oriented problems. It's a demonstration that ARM provides a significant cost savings in up-front hardware and ongoing overhead (electricity) vs. the previous supercomputing architectures built with GPU and x86 hardware.

      The key is really custom silicon as opposed to ARM vs x86. The whole thing about the Fujitsu machines is they have a very fast, very wide FPU on the same die as a very fast very low latency interconnect (Tofu), with the CPU core there to shuffle data between the two. This gives Fujistu a latency advantage over other supercomputer designs which is a good part of why they do so well at utilization and efficiency. Sueprcomputers live or die on the interconnect.

      The older SPARC machines were similarly efficient compared to the contemporary x86 and GPU machines for much the same reason.

      • by AmiMoJo ( 196126 )

        +1 informative. Some of the work it did before completion on droplet spread was very interesting, and kind of disgusting.

        https://youtu.be/GAvO_QdO9eM [youtu.be]

        There was another which I have temporarily misplaced showing a desk with a partition divider that is supposed to protect the users from each other, and the stream of particles going over it.

      • by Junta ( 36770 )

        Adding another fun fact, Intel's acquisition of Qlogic's infiniband to create omnipath at one point had a goal of having a nicely integrated fabric controller on the processor package very nicely integrated with the CPU, to be able to compete with this sort of strategy.

        Then they backed off to it being on the CPU package, but connected through mundane PCIe, and released that (the worst of both worlds, no particular latency/bandwidth benefit, but having to share power and thermal budget with a pretty intense

        • Yeah, Infiniband is pretty much dead, only living on in a zombie form that tunnels through Ethernet and some rusty old legacy hardware. Good riddance too, now would it be so kind as to die the rest of the way.

          • by Junta ( 36770 )

            Actually, nVidia is still big on infiniband.

            It's still much easier to build very big low latency networks with infiniband than ethernet.

            Of course at a core signaling level IB is now very close to ethernet, but the frame format and handling are very different.

            It's a mostly source routed protocol (switches can select alternate equivalent hops due to congestion that the source was not privy to) that doesn't care if you loop so you can make torus, fat tree, or other topologies without much of a care (though you

            • Nvidia is only big on infiniband because it bought Mellanox. There is exactly one Infiniband vendor still standing: Nvidia/Mellanox. Global infiniband revenue is roughly $200 million. Less than two percent of Nvidia's revenue. Less than three percent of the global ethernet cable market. The industrial ethernet market alone is $35 billion. I think the correct adjective that applies to infiniband is "dead".

              • by Junta ( 36770 )

                Yes, they are 'only' big on infiniband because they spent the money to buy the infiniband market. I know, they also represented a cheap yet respected merchant of ethernet silicon compared to Marvell or Broadcom.

                Infiniband revenue has been growing over time, rather than declining. It's niche to be sure but it's no more niche than it used to be, and the revenue story suggests the trend is not away. Sure it wasn't the cross-vendor story people were hoping for, but as a technology it is persisting.

                Now will n

                • Nvidia did not buy Mellanox to get Infiniband, they wanted the Ethernet business, many times the size. While It's hard to know the future, the past gives provides guidance: commodity invariably drives niche extinct. The arguments in favor of Infiniband are basically just: 1) latency. But look here. [orthogone.ca] 196 ns round trip. So much for Infiniband's latency argument. Amazing what a vibrant IP core ecosystem can do. In truth, ethernet latency has been mostly in the software stack, not the hardware. At the transport

      • The key is really custom silicon as opposed to ARM vs x86.

        It's not custom unless you want to say that every ARM SoC is custom. Fugaku uses 512 bit SVE [wikipedia.org] from ARM.

        • It's not custom

          Except it really is custom.

          unless you want to say that every ARM SoC is custom.

          One of the big advantages of ARM is that you can in fact have custom silicon. People can do anything from chuck an off the shelf core on the same die as some off the shelf blocs for other things, or mae big modificiations lie this CPU.

          Fugaku uses 512 bit SVE from ARM.

          "from" ARM sure. So far it's the only processor that uses those instructions. And they use them to drive their own floating point unit which is an a

    • Also missing from the summary:

      Fujitsu A64FX ARM-based CPUs, with 512-bit vector extensions and a custom torus interconnect
      HBM2 Memory
      ~1.5TB NVMe storage
      Running custom Linux kernel

      https://en.wikipedia.org/wiki/... [wikipedia.org]

  • i wonder how much hashrate you could get with this thing running eth

He who steps on others to reach the top has good balance.

Working...