The advertisements above do not necessarily reflect the views of this blog, its authors, or host.

IBM Computer FLOPS Big Time

IBM has outdone itself again by creating (again) the world’s fastest supercomputer!

Normally this would deserve page 2 or 3, but in Bob’s world this warrants page 1 (above the fold) because it is the first computer to enter the long sought after realm of PetaScale computing. It is capable of 1.026 PetaFlops which means it can do more than 1 quadrillion (thousand trillion) calculations per second. Your speedy desktop computer is an abacus by comparison with maybe 5 to 10 billion. You may see the term FLOPS a lot in this context. This stands for FLoating Point Operations Per Second which, as you probably guessed, is roughly equivalent to how many instructions per second it can carry out.

It is called Roadrunner and was developed by the Department of Energy–slash–US Military’s Los Alamos National Laboratory and specialists from IBM. I don’t care much for that name of this supercomputer though. I’d like to officially suggest:

The Great Hyperlobic Omni-Cognate Electron Wrangler

Just a thought.

The standard test that is done to determine supercomputer speed is called the Linpack benchmark. It consists of solving linear equations involving about 2 million equations and the same number of unknowns. The test was carried out at IBM’s Poughkeepsie plant in New York. That’s under an hour away from me. I wish I knew that earlier. We could have planned a heist. Maybe I’d be able to actually play that new Conan game without any lag if I had petaflops to work with.

To be fair, some researchers claim to have already achieved this level of performance but their machines are so specialized that they can’t run the benchmark software.

Anywho…the architecture is pretty impressive.

It breaks the recent trend of adding more and more distinct processors. The now third place supercomputer has 65,000 processors and 2nd place has almost a quarter million. Roadrunner has just 20,000.

7,000 of them are comparatively ordinary Opteron processors from AMD (Advanced Micro Devices) which are used in many corporate servers. The remaining 13,000 (with 8 cores each) are upgraded versions of IBM’s Cell chips used in Sony’s Playstation 3 video game console. I believe this Hybrid technique is a first.

Roadrunner is financially and physically intimidating as well. For 133 million dollars you get 288 refrigerator-sized cases and that needs 3 megawatts to run. This is comparable to the power used by a large suburban shopping center.

Now it’s time to address the inevitable “Big Deal” response I get from time to time.

This is a big deal. This level of computing is not just the odometer effect like your car hitting 100,000 miles or going from the year 1999 to 2000. PetaScale computing has the potential to fundamentally alter science and engineering.

According to John Morrison, leader of the high-performance computing division at the Los Alamos National Laboratory (LANL) in New Mexico, where it will be installed:

“Roadrunner will enable us to tackle problems we couldn’t tackle before, we’ll be able to run a different level of problems…to do calculations that we wouldn’t even consider before.”

“We’ve proved some skeptics wrong,” says Michael Anastasio, director of the Los Alamos lab. “This gives us a window into a whole new way of computing. We can look at phenomena we have never seen before.”

For a real-world example of how this could help, consider medical imaging. Analyzing a CT-Scan is one of the most powerful diagnostic tools available today but it is computer intensive. Currently at 50 trillion FLOPS, it can take up to two hours to analyze a single CT scan. At a Petaflop, we could do that in real time.

So what’s the next big hurdle in terms of raw speed? The next order of magnitude milestone will be the exaflop computer. This will be able to crunch 1 quintillion calculations per second (a million trillion). Maybe when this happens people will finally know why I have “exabyte” in one of my email addresses :-)

Like the petaflop, the exaflop will give us unprecedented resolution for simulating reality.

Take Genomics Research for example. In order to model a cell’s metabolic and signaling network, it would take 6 days on a machine like Roadrunner. An Exaflop based computer should be able do it in close to real time. It may be a while though before I blog about IBM’s first exaflop computer. 2019 is the best guess at this point.

Wanna take another 3 orders of magnitude step? What about the Zettaflop? This 10 to the 21 operations per second or 1 sextillion. What a great number that is. Definitely in the top five for me along with 13, 0, 666, and Grahams Number (look it up).

What might this accomplish? It’s hard to say what layers of reality this level of computing might reveal but NASA theorizes it could accomplish “full weather modeling” (whatever that really means) which could cover a two week period with incredible accuracy. My guess is it could predict the weather two weeks in the future with the accuracy we expect now from a two-day forecast. This might not happen until 2030 or so.

How far can we go? What are the physical limits to computing as we understand them?

10 to the 51 operations per second has been calculated as the upper limit. Those pesky Quantum rules won’t let numbers be crunched any faster, apparently. The designation for that speed is not sexy like petaflop or exaflop is. Unfortunately it might be called a Pepta-PP-Flop supercomputer. Doesn’t roll off the tongue, does it?  Maybe by the time that computer arrives, someone will remember this post and call it The Great Hyperlobic Omni-Cognate Electron Wrangler.

16 comments to IBM Computer FLOPS Big Time

  • sago

    Interestingly, the peta-scale computer is also roughly the size needed to simulate the number of neurons in the human brain as an artificial neural network in roughly real time (1PFLOP/s is about 0.25% real time – so within 1 order of magnitude).

    It means that we have built the first brain scale computer.

    Obviously is isn’t going to be used for AI research, but it gives you a sense of how far we’ve come in computing machinery.

  • spurge

    That is very cool. I have a soft spot for Big Blue.

    I grew up in Poughkeepsie and even worked as a temp building main frames there.

    Many people in my family have and still do work there.

    I will have to see if my relative who works in chip design worked on the ones for the new super computer.

    Maybe I can get you a tour?

  • larry coon

    Good news….Roadrunner just scored a 5.5 on the CPU performance meter in Windows Vista.

  • John Powell

    The Singularity draws nearer…

    I for one welcome our post-human overlords.

  • I’ll put in a good word for you John 1.0

  • Belgarath

    Bob,

    I believe that’s supposed to be:

    Great Hyperlobic Omni-Cognate Neutron Wrangler.

    Deep Thought’s thoughts on this machine:

    The Great Hyperlobic Omnicognate Neutron-Wrangler can talk all four legs off an Arcturan Mega-Donkey but only I can persuade it to go for a walk afterwards. Molest me not, with this, pocket calculator stuff!

  • Traveler

    I don’t think that 10^51 absolute limit applies in this situation. Remember that we are talking about computers with lots of processors in them. No one piece of the computer gets close to the published FLOP rate by itself.

    As for getting 2 week weather forecasts that are as accurate as today’s 2 day forecasts — I’d settle for getting an accurate 2 day forecast.

  • spurge

    I seem to have a comment stuck in moderation.

    I was wondering why?

  • petrucio

    Traveler,

    10^51 is the limit of computation you can get out of 1Kg of matter (IIRC), it does not matter what architecture you are using.

    My computer has 2 processors, but it is still one computer. That distinction does not apply. Multiprocessors is just a change in architecture paradigm, just like using pipelines was, and no one complained about that.

  • petrucio

    Hummmm, Kurzweil predicted petaflops in 2007, missed by little. But that was a recent prediction, so not fair. Computational power needed to simulate a human brain is predicted in the next 10 years, and then 10 more years for all that muscle to be available for under $1000 (I think).

    So it looks like about 20 years before your computers outsmarts you. Neat stuff.

    So it looks like the rhythm really IS going to get you! I mean – the Singularity really IS near!

  • Martinus

    “Maybe by the time that computer arrives, someone will remember this post and call it The Great Hyperlobic Omni-Cognate Electron Wrangler.”

    Or they could go with Douglas Adams and call it The Great Hyperlobic Omni-Cognate Neutron Wrangler.

    I’ve been a fan of the HHGTTG since the first radio broadcast, and I knew something seemed wrong when I first read that line…

  • So I went and looked up Grahams Number. Go read the Wikipedia first paragraph entry then tell me it wasn’t written by Douglas Adams.

  • I bet technicians still play Solitaire on it in their coffee break.

  • Code run on the machine mimics brain mechanisms underlying human sight.

    http://www.lanl.gov/news/index.php/fuseaction/home.story/story_id/13602

  • DLC

    Great Hyperlobic Omni-Cognate Electron Wrangler ??
    GHOCEW ? don’t get it. But then, you knew that DLC stands for Digital Logic Circuits, right ?

    The Reptillian Overlords will be wiped out by the Digital Overlords.

    You shall be assimilated. resistance is Ohms.
    Or futile, or something.

  • Thanks for the comments everyone. I was beginning to think only my mom was reading this.

    Regarding the fanciful name that I borrowed from Douglas Adams. I loved it when I first came across it years ago but “Neutron Wrangler” wasn’t right for this post because if Roadrunner wrangles anything it must be electrons. That’s why I made that change.

Leave a Reply