back to article 5.34 petaflop SGI monster forecast at US National Center for Atmospheric Research

SGI and partner DataDirect Networks will spend 2016 beating a path to America's National Center for Atmospheric Research (NCAR), which named them to deliver its next system in 2017. The iron will replace its 1.5 petaflop Yellowstone with an Intel Xeon-based 5.34 petaflop thumper to improve NCAR's climate and weather modelling …

  1. Anonymous Coward
    Anonymous Coward

    Since there isn't one of the climate change models that actually predict correctly how is this going to improve anything?

    1. This post has been deleted by its author

      1. Ol' Grumpy

        "Apparently, it runs Crysis really well :-)"

        But still not at full detail ;)

  2. Tom 7

    3 billion calculations per second, per watt.

    That's about 10% of the parallella which does 64GFlops in 2w.

  3. Pirate Dave Silver badge


    So, I wonder what they're going to do with all those horribly outdated and crappy nodes in Yellowstone once Cheyenne gets fully up and running? I mean, Xeons from 2012 are pretty much worthless for anything useful, except maybe as boat anchors. So, as a good American, I'll offer some space in my server room to store some of those old racks full or useless junk for a few years. It's my patriotic duty.

    and from the sister-site article about the new system:

    "The climate code is over 1.5 million lines of code and there are over 100,000 IF-THEN statements, and those kinds of things don’t do as well in GPU-type architectures."

    "IF-THEN" statements? The climate code was written in VB6? Yeah, I can't imagine VB translates well to GPU architectures.

    1. robert lindsay

      Re: Yellowstone

      fortran, lots and lots of fortran.

  4. robert lindsay

    Research Box only

    NCAR doesnt do operational work. This is a research box only. (and vastly more powerful than NOAA's operational machines)

  5. Anonymous Coward
    Anonymous Coward


    5.4*1000*1000/145000 = 37:

    So at 37GF per core, it has got to be a Broadwell xeon in 2017 (2.3GHz ish)

    2.3*4*2*2 = 36.8 GF.

    (clk)*(vecwidth)*(fma)*(dual issue)

    and not Skylake...

    2.3*8*2*2 = 73.6 GF.

    Interesting choice - was it price, poor vectorisation, or schedule?

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like