back to article Google reaches into own silicon brain to slash electricity bill

Google has worked out how to save as much as 20 percent of its data-center electricity bill by reaching deep into the guts of its infrastructure and fiddling with the feverish silicon brains of its chips. In a paper to be presented next week at the ISCA 2014 computer architecture conference entitled "Towards Energy …

COMMENTS

This topic is closed for new posts.
  1. Anonymous Coward
    Anonymous Coward

    All good stuff

    For commercial customers your prices are determined by four things:

    How much you use

    When you use it

    The maximum you use in any half hour

    The grid connection capacity

    So a 35% power saving won't alter maximum demand or grid connection capacity needs as these are set by the data centre running at full chat. As the power savings will be maximised at lower utilisations (so off peak) the unit cost savings will be lower than Google's average unit rate as well.

    At a guess this is a 5-7% cost saving on the total power bill rather than the claimed 20%. Still worth working for, but nothing to get in a lather about.

    1. John Robson Silver badge

      Re: All good stuff

      Unless you produce yourowm power...

      1. Anonymous Coward
        Anonymous Coward

        Re: All good stuff

        "Unless you produce yourowm power..."

        Actually no, it's just the same. If you produce your own power at any scale then you've got exactly the same issue that your capacity investment is fixed, but your power demand varies. That's why wholesale and B2B power prices vary by day and by season, because meeting peak demands means more lightly utilised plant.

        What Google are doing (in the article) is reducing off peak demand, and that's the least useful form of power reduction whether you generate in house or buy from the grid; What you really want to do is reduce your peak load, or to reschedule your peak load so that it is different from the peak grid demand. To an extent they can already do load shifting across time zones where the data connections permit that, but as far as I know there's not sufficient cable capacity to run all of continent A's data processing on continent B during continent B's dark hours when power demand is minimal and prices low. And as we now know, you most certainly can't trust continent B.

      2. Shady

        Re: All good stuff

        Unless you produce you owm power?

        What have Smeg and the Heads got to do with it?

        1. king of foo

          Re: All good stuff

          It's cold outside...

          Etc?

    2. Robert Helpmann??
      Childcatcher

      Re: All good stuff

      At a guess this is a 5-7% cost saving on the total power bill rather than the claimed 20%. Still worth working for, but nothing to get in a lather about.

      Let's do the math... back of the envelope stuff... 5-7% - let's say 6% - multiplied by Google = lots and lots of lots. I would get in a lather if they offered that to me! More seriously, telling investors that they saved a significant amount of money will only drive stock upward.

      1. Tom 38

        Re: All good stuff

        Especially as in a DC, power is your biggest cost. If your devices use less power, then they generate less heat. If they generate less heat, you can pack more of them in per rack. If you can pack more in per rack, you can have more devices in the DC period, and the cost of hosting is reduced.

        I'm sure google have well specified DCs, but we often can't fill our rack because our DC provider can't* sell us more power, because they are near their cooling limit. If we used less power, we could have more per-rack, and we would need less racks/have more servers.

        * Of course they can give us more current, but it is exponentially expensive, to the point where filling a rack (using say 22A) is almost as expensive as taking another whole 13A rack.

        1. Gordon 10

          Re: All good stuff

          Interesting point. I wonder if the article includes the power reduction for data center cooling into the equation - although you cant run the AirCon on the same kind of latency - there must be an aggregate effect thats worth noting.

          Anyone know the ratio of Rack power to DC power in a data center?

  2. Destroy All Monsters Silver badge
    Thumb Up

    I hope we will see this in interplanetary robot probes.

    Someone will just have to backronymize "GALACTICA" for maximum effect.

    1. Ben Bonsall

      god awful lame acronym chosen through invoking creative apathy.

    2. Tiny Iota
      Coat

      Greetings Alien Lifeform. Always Capitulate Totally If Commentards Attack!

      or just say something snarky and downvote.

  3. John Smith 19 Gold badge
    Meh

    So it looks like some kind of simulated annealing optimization process run on the load

    Just a wild guess.

    Too bad no one's prepared to build a clockless intel compatible processor.

    1. Destroy All Monsters Silver badge
      Holmes

      Re: So it looks like some kind of simulated annealing optimization process run on the load

      Too bad no one's prepared to build a clockless intel compatible processor.

      Wasn't it stopped because using that many transistors for the clocklessicity wasn't (yet) worth it?

      Well, when it becomes economically interesting, it will be back on the board. Then comment fiends will complain "what an old idea" etc. etc.

  4. Matt Bryant Silver badge
    Meh

    Meh.

    So, Google have their own trays, in their hand-designed DCs, and they have the problem that their previous energy control amounted to two States - core on or cores off. So, now they have sensors that allow them to match demand by tuning kit up and down. Yay, go Google! Yawn.

    This illustrates exactly the problem with DIY solutions over the integrated solutions you get from Dell, hp, Fujitsu or IBM - you have to make the clever stuff yourself. Intelligent system management that looks at what the systems are doing and adjusts power (and intelligent localised cooling, something I assume Google are still working on) have been around on COTS gear for quite a while now. And it's not just the hardware vendors, VMware have had the ability to set power policies in vSphere for a while too. So, yes, Google are clever boys and girls, but this seems to me to be just catch-up with commercial capabilities.

    1. Destroy All Monsters Silver badge
      Thumb Down

      Re: Meh.

      > "Integrated solutions" you get from Dell, hp, Fujitsu or IBM

      > "DIY" solutions

      So, the companies that actually are at the coalface of large-scale deployment work are setting up "DIY" solutions (the horror!) while big-name premium box-shifting brands (most of them circling the drain) that don't even make the map in processor count have it already all figured out?

      That's brand cargo cultism at its worst.

      1. Matt Bryant Silver badge
        Facepalm

        Re: Destroyed All Braincells Re: Meh.

        ".....the companies that actually are at the coalface of large-scale deployment work......" Whilst Google certainly are at the coalface, you do not show that they are ahead in developing their solution over commercial offerings. We were making use if similar tech in production DCs waaaay back when Carly Fiorina was still selling us her Adaptive Entreprise kit (Proliant G6 days?), and I'm pretty sure then or then abouts with IBM Power gear too. Google are simply trying to cut out cost by building their infrastructure themselves, but they would seem to be following in the footsteps of others rather than blazing a brand new path.

        1. JackClark

          Re: Destroyed All Braincells Meh.

          Hiya, You may be mistaken - a reader got in touch recently to tell me that IBM recently demoed a beta of Platform Symphony LSF using RAPL for power management, so I'm not sure Google are that far behind (or even behind at all, given that this work at Google was likely done some months/years ago). Also, the distributed controller scheme strikes me as novel as well.

  5. Anonymous Coward
    Anonymous Coward

    Hmmm...

    So they were not doing this before, like Azure and AWS...?

    Google is way behind the curve.

  6. Sureo

    Why don't they do something useful with their spare cycles like distributed computing (BOINC)? That's what I do on my computers at home.

This topic is closed for new posts.

Other stories you might like