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

 



Forgot your password?
typodupeerror
×
Japan Hardware Technology

Japan's ARM-Based Supercomputer Leads World In Top500 List; Exascale Expected In 2021 (techtarget.com) 25

dcblogs writes: Japan's Fugaku ARM-based supercomputer is the world's most powerful in the latest Top500 list, setting a world record of 442 petaflops. But this was otherwise an unremarkable year for supercomputers, with a "flattening performance curve," said Jack Dongarra, one of the academics behind the twice-a-year ranking and director of the Innovative Computing Laboratory at the University of Tennessee. This is a result of Moore's Law slowing down as well as a slowdown in the replacement of older systems, he said. But the U.S. is set to deliver an exascale system -- 1,000 petaflops -- next year and China as well. Meanwhile, the EU has a 550 petaflop system in development in Finland. "On the Top500 list, the second-ranked system was IBM Power Systems at nearly 149 petaflops using its Power9 CPUs and Nvidia Tesla GPUs. It is at the Oak Ridge National Lab in Tennessee," adds TechTarget.

"Third place went to Sierra supercomputer, which also uses Power9 and Nvidia GPUs, at about 95 petaflops. It is at Lawrence Livermore National Laboratory in Livermore, Calif."
This discussion has been archived. No new comments can be posted.

Japan's ARM-Based Supercomputer Leads World In Top500 List; Exascale Expected In 2021

Comments Filter:
  • by grumpy_old_grandpa ( 2634187 ) on Tuesday November 17, 2020 @02:46AM (#60733404)

    I would think the ARM-based architecture is the real news here. I haven't paid too close attention over the years, but reckon the Top 500 has been mostly about x86 and PowerPC.

    A link to an article which discusses that rather than going into musings about Moore's law and the future of exa-scale would be interesting.

    • by An Ominous Cow Erred ( 28892 ) on Tuesday November 17, 2020 @03:30AM (#60733476)

      Pretty much. We're hitting the theoretical limits of silicon now (not the theoretical engineering limits, the theoretical PHYSICAL limits). This also means we're basically hitting the physical size limits of a 2D processor of any type. To advance we're gonna need a fundamental improvement in materials or physical layout. This means either a way to create 3D processors (TRUE 3D, not just stacking a few extra layers), or finding a new material that allows higher clock speeds either electric or photonic. ...and of course we have to figure out how to cool these new monstrosities.

      • by nojayuk ( 567177 )

        ...and of course we have to figure out how to cool these new monstrosities.

        This problem was recognised back in the 1980s when the ever-increasing density of computing silicon (sub-1000nm layout processes!) led to the acronym HSG being coined. The Hairy Smoking Golfball was to be the new unit of computing -- the HSG was covered in wires to get signals in and out of the core (16 bit and even, gasp, 32 bit data buses!), it was smoking from the heat dissipation (a hundred watts and more!) and it couldn't be mu

    • by serviscope_minor ( 664417 ) on Tuesday November 17, 2020 @05:05AM (#60733640) Journal

      I would think the ARM-based architecture is the real news here. I haven't paid too close attention over the years, but reckon the Top 500 has been mostly about x86 and PowerPC.

      It is a bit but not how you think. ARM isn't displacing x86 or PPC, it's displacing SPARC. Fujitsu were the last supercomputer vendor of UltraSparc chips. Basically the SPARC was a relatively slow, simple core they could use a ton of to connect their very fast low latency on die interconnect to their very fast very wide floating point unit (and do all the data transfers, marshalling etc). They decided making entirely their own core just for this was not of sufficient benefit, so they switched to ARM. They're still using their own interconnect and their own FPU however.

      • by Bert64 ( 520050 )

        Well they used to collaborate with sun on sparc, but oracle have pretty much given up on sparc so it makes sense for fujitsu to move to something else.

        A lot of supercomputers are using slow general purpose cores to control a large number of gpus or other specialist hardware, or using a large number of relatively low clocked cpus because higher clocked ones would just stall waiting for memory given the machine's intended workload.

      • by _merlin ( 160982 ) on Tuesday November 17, 2020 @09:40AM (#60734144) Homepage Journal

        Fujitsu were the last supercomputer vendor of UltraSparc chips.

        Fujitsu never used Sun UltraSPARC CPUs. They used their own SPARC64 CPUs. UltraSPARC and SPARC64 are both implementations of SPARCv9, but they were developed separately.

        They decided making entirely their own core just for this was not of sufficient benefit, so they switched to ARM.

        It's still their own core. It isn't an off-the-shelf core from ARM, Qualcomm or anyone else. It's a pure AArch64 implementation with no 32-bit ARM backwards compatibility, and no NEON SIMD unit. It has their own scalable vector unit instead. They switched to AArch64 because it's a cleaner, more modern architecture than SPARCv9, and they figured they could get better performance/watt out of it. Distributing power and removing heat is a big issue in supercomputers.

        • by Misagon ( 1135 )

          ... no NEON SIMD unit. It has their own scalable vector unit instead.

          They might have developed their own scalable-vector unit, but it is using the ARM SVE ISA, which is a superset of AArch64's regular SIMD ISA and uses the same register file -- with longer registers.

          Support for SVE is in GCC and clang/LLVM (including ARM's fork of the latter) -- which Fujitsu wouldn't have been able to benefit from if they had been using their own ISA. Compiler-writers have been using instruction set emulators for testing.

          • SVE is their own instruction set. It isn't a superset of NEON. They published the specification and initial GCC support themselves. There currently aren't any SVE implementations from any other vendors. (The integer and flow control instructions are all identical to AArch64 and use the same registers, but that's not where the special supercomputer sauce lives.)

    • I would think the ARM-based architecture is the real news here. I haven't paid too close attention over the years, but reckon the Top 500 has been mostly about x86 and PowerPC.

      POWER, not PowerPC. PPC has been all but abandoned, it lives on in embedded but only in small volume compared to ARM.

      (And amd64, not x86...)

    • It's just a housekeeper.

      You could have huge differences in the performance and power efficiency of the housekeeper core and it would matter naught to the performance of the supercomputer. The silicon doing the real work has got nothing to do with ARM.

  • Bet they still have to wait for Windows to update.
  • The Top 500 supercomputers list proves it only too well. How right was the Monkey Boy.
    • Are you joking, right? Even gaming is a Linux thing today...
      • He's either trolling, or dumb as shit.

        Linux underpins the world's communications infrastructure today.

        It has absolutely dominated the server space.

        The majority of the world's smartphones run on Linux.

        And more directly on-topic, Linux positively dominates the top500 [itsfoss.com].

        Nobody who knows anything thinks that Linux is done for.

        • The majority of the world's smartphones run on Linux.

          And the rest of the majority of those phones runs on a BSD-based OS. Microsoft tried too hard to cram the experience of the Windows desktop on the small screen of a phone.

          So basically, nearly all modern phones run on a Unix-based (iOS) or Unix-like (Android) variant of some kind.

          And with most software and services going on the Web because it's easier to make it subscription-based and make sure everyone is always up-to-date, it wouldn't surprise me to see

  • Nope, I guess not. Well, that figures :-) The godz alone know what goes on in the basements of some of those buildings at Fort Meade.

"Ada is PL/I trying to be Smalltalk. -- Codoso diBlini

Working...