back to article Who needs a supercomputer when you can get a couple of petaflops on AWS?

Descartes Labs, an outfit that analyses big data, has managed to nab the 136th spot on the top 500 list of the world's fastest publicly known supercomputers – with $5,000 and an Amazon Web Services (AWS) account. The AWS "supercomputer" has 41,472 cores, 157,824 GB RAM, and achieved 1,926.4 TFlop/s using the LINPACK benchmark …

  1. Aladdin Sane

    Something, something, something, Crysis.

    1. Anonymous Coward
      Anonymous Coward

      "Another issue is performance. "AWS may run LINPACK well, which is notoriously easy. I/O performance sucks on AWS. A more HPC-like infrastructure has RDMA (Remote Direct Memory Access) networks. "

      Like Azure for instance.

      1. Enki

        Or AWS on some instances (https://perspectives.mvdirona.com/2019/02/aws-nitro-system/). I wouldn't be surprised if Google can on some cases as well.

    2. Grooke

      Obligatory smbc:

      https://www.smbc-comics.com/comic/2011-02-17

  2. John Robson Silver badge

    3x cheaper...

    But only if you have enough tasks to keep it busy...

    1. Yet Another Anonymous coward Silver badge

      Re: 3x cheaper...

      But for all those tasks where you just need a supercomputer to model a galaxy collision or nuclear weapon detonation once before going back to your day job - AWS is perfect.

      It's why the rental model for aircraft carriers is so popular, you only need them when you want to invade somewhere so you don't want to pay for one 24x7

      1. Stevie Silver badge

        Re: 3x cheaper...

        Aircraft carrier plus crew, of course. Can't ask a crew to put to sea in a vessel they have never sailed, maintained or fought before.

        1. Yet Another Anonymous coward Silver badge

          Re: 3x cheaper...

          Obliviously you keep a highly trained and proficient crew on Zero Hour contracts

          1. Robert Sneddon

            Re: 3x cheaper...

            Obviously you keep a highly trained and proficient crew on Zero Hour contracts

            That's how the Royal Navy used to operate to man their ships -- press-gangs would go around a port's taverns and knock some landlubbers over the head and when they woke up they were aboard ship and out of sight of land. Nowadays the DWP could do the same thing...

            1. Arthur the cat Silver badge

              Re: 3x cheaper...

              Nowadays the DWP could do the same thing...

              The DWP would be more likely to knock themselves over the head and then dump themselves in the sea without a ship. Another fine IDS notion.

            2. Anonymous Coward
              Anonymous Coward

              Re: 3x cheaper...

              Most of the people pressed were competent sailors, I believe. Hitting people to get them to work the capstan is all very well but it's inefficient, and the RN only used to rely on non-sailors as a last resort.

              Sometimes warships would simply stop merchant ships returning to port and steal their crews.

              1. Youngone Silver badge

                Re: 3x cheaper...

                Sometimes warships would simply stop merchant ships returning to port and steal their crews.

                One of the reasons for the War of 1812 I think.

                Still, that ended with His Majesty's brave troops burning the White House down, so a good result.

                1. Anonymous Coward
                  Anonymous Coward

                  Re: 3x cheaper...

                  It started with the US burning down Toronto - an even better result (Vancouver resident...)

      2. JLV

        Re: 3x cheaper...

        There’s a joke there somewhere, at the expense of UK taxpayers, when US F-35 pilots fly on catapult and plane-less UK helicopter carriers:

        https://www.military.com/daily-news/2019/05/07/us-marine-f-35-squadron-will-deploy-british-aircraft-carrier-2021.html

    2. Bronek Kozicki

      Re: 3x cheaper...

      That's why hybrid cloud makes the most sense - buy the infrastructure for the baseload and offload everything else. That's what OpenStack is meant to do best.

  3. LeahroyNake

    Approximately

    "We were granted access to a group of nodes in the AWS US-East 1 region for approximately $5,000 charged to the company credit card."

    I hope they filled out their expenses claim form correctly for that one!

    1. Anonymous Coward
      Anonymous Coward

      Re: Approximately

      Meh - depends on who is doing the asking. I have seen similar numbers signed off for "knicker elastic" on expenses claims!

      1. Bronek Kozicki
        Paris Hilton

        Re: Approximately

        Do you work in parliament?

      2. The Nazz

        Re: Approximately

        Affectionately known as "loose knicker elastic" at one company i worked for, whose staff (the lucky ones, maybe) were able to overrun the generous entertainment budget. Tbf, they did have a point of need in that the Europeans spent a hell of a lot more entertaining them than our company did them.

  4. Robert Sneddon

    Wee Archie

    Edinburgh Uni's High Performance Computing Centre has an ageing supercomputer named Archer (9,480 12-core Xeons). Back in 2014, just after it was inaugurated it ranked 19th in the TOP500 supercomputer list, today it's down about 250 or so. They've also got Wee Archie, a 16-node Raspberry Pi supercomputer they use for educational use and lectures. It costs about 2000 quid and instructions on how to build your own version are available from Edinburgh University's HPCC website.

    https://www.epcc.ed.ac.uk/discover-and-learn/resources-and-activities/what-is-a-supercomputer/wee-archie

    Apparently in terms of performance Wee Archie would have been one of the top 100 supercomputers back in the 1990s...

    1. Anonymous Coward
      Anonymous Coward

      Re: Wee Archie

      Back in 1990s we had a Cray-2 and a Cray-YMP (in a building guarded by some unsympathetic men with machine guns.) They were about 2 GFlops each.

      A Raspberry Pi 3 is about 5GFlops on CPU alone.

      The PC graphics card i'm using for this job is allegedly 15,000 GFlops

      1. Robert Sneddon

        Re: Wee Archie

        The "build your own Wee Archie" webpage ran LINPACK on a five-machine cluster and got 2.4Gflops/s. A single Pi 3B was about 270Mflops/s.

        1. Anonymous Coward
          Anonymous Coward

          Re: Wee Archie

          Sorry missed a decimal point, yes 3B+ should be around 0.4-0.5 GFlop

      2. vtcodger Silver badge

        Re: Wee Archie

        A Raspberry Pi 3 is about 5GFlops on CPU alone.

        And, I'm guessing that a typical modern web page --- amazon.com for example -- will bring it to its knees for several minutes while trying to download and render the incredible amount of junk that appear to be essential to my contemplated purchase of 100 index cards.

        (I AM impressed with the capability of modern CPUs. The stuff running on them ... not so much)

  5. Anonymous Coward
    Anonymous Coward

    Re: AWS and I/O

    Maybe AWS's I/O performance wouldn't stink so bad if there were fewer Java tracebacks flying around their network. I wonder if anyone's been able to determine just how much of their network traffic is Java error messages being sent between AWS systems. Based on what I saw when I last worked on AWS-based systems, it can't be negligible.

  6. Anonymous Coward
    Anonymous Coward

    "The AWS "supercomputer" has 41,472 cores, 157,824 GB RAM"

    Yes but is it powerful enough to run Windows Vista?

    1. Anonymous Coward
      Anonymous Coward

      Re: "The AWS "supercomputer" has 41,472 cores, 157,824 GB RAM"

      We tried! But only 3 cores were detected, 4gb of ram accessible, and it blue screened after 12 minutes of runtime; we gave up. Someone suggested trying Windows 10, but he was mysteriously pushed down a lift shaft.

  7. hoola Silver badge

    The entire point of an HPC cluster is to manage it such that it is working at capacity, that is what it is for and generally there is more workload available than there is capacity. This is just willy-waving by someone to try and make AWS look good.

    Vested interests?

  8. MachDiamond Silver badge

    IPSec

    If you are doing work that needs a supercomputer, you probably also need tight security. Security is a word I don't associate with AWS. There are centers with various levels of computing performance where the hardware is not online and you are in the room with the beasties so you can assure yourself that your program and data are dumped when you finish to a much higher level of certainty.

  9. inquisitive2014

    $5,000/3 = $1,667 for a Supercomputer at Cambridge.

    I am skeptical of the claim that Cambridge can do it for 3X less. I suspect that they are forgetting to add in a lot of often overlooked costs like rent on the Datacenter, all staff costs, redundancy etc. While it is true that it is possible to run IT systems for less than the major cloud providers it is extremely difficult to do it to the same standards. My experiences in Australia with the largest corporations and government IT systems has shown me that less than 5% of them are delivering to the standards of the major cloud providers regardless of cost.

    Additionally I dont know of any Supercomputer project that has only spent $5K on infrastructure.

    The case for on-premise IT is getting harder and harder to justify

    1. Anonymous Coward
      Anonymous Coward

      Re: $5,000/3 = $1,667 for a Supercomputer at Cambridge.

      At my institution (also in Australia), we do it for ~5x less, at scale :) If we didn't, our department would have been outsourced by now.

      I have almost a thousand projects running on the slurm based cluster I built and operate. We service hundreds of users daily, and most of them would be far less efficient having to build and maintain software themselves, etc.

      It's kinda cool that you think that's not possible. Gratifying tbh.

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

Biting the hand that feeds IT © 1998–2021