4 ms·

Take the post's estimate at face value: 1 MW to run AlphaGo. Residential electricity costs on the order of $0.10/kWhr, or $10/hour for 1 MW. If the same machine

Take the post's estimate at face value: 1 MW to run AlphaGo. Residential electricity costs on the order of $0.10/kWhr, or $10/hour for 1 MW. If the same machine were a better programmer than you, would you say "pshaw, come back when it beats me at power parity"? OK, the total cost of operation will be substantially higher (though it probably runs on cheaper electricity), maybe even approaching your total cost to your employer. This does bring up a point I've wondered about: why haven't we yet found a productive way to spend lots of computing power on making us better programmers? Usually the most cycles are spent on compiling and on test runs, costing only a tiny fraction of our hourly rates. (Maybe that compute budget for software development looks at least somewhat different at Google and some other places.)

jacquesm 11y ago
$10 will buy you 100KWh, not 1MWh, you're off by a factor of 10. > why haven't we yet found a productive way to spend lots of computing power on making us better programmers? Usually the most cycles are spent on compiling and on test runs, costing only a tiny fraction of our hourly rates. The Go language (ironically) addresses some of this by putting programmer productivity very high up in their feature list by focusing on compiler speed to the point where some of the differences between compiled and interpreted languages disappear.
- DonaldFisk 11y ago
  Turbo Pascal also had very fast compilation times (no more than a few seconds for large programs) 30 years ago, and it worked well then. But you don't need to compile the whole program, just the functions you have changed. With Lisp incremental compilation is often used, and it is almost instantaneous - just point the mouse at the function you want to recompile in EMACS, hit the load "chord", and it's ready to run in the listener. You're not even aware the compiler is there. Of course, there's a lot more to programmer productivity than compilation speed, as Fred Brooks pointed out.
- abecedarius 11y ago
  Oops, thanks for the correction. OTOH zamalek elsewhere in these comments thinks 1MW is a considerable overestimate.
  - jacquesm 11y ago
    The 1MW figure was based on what I could get out of the available articles, if anybody has a better estimate backed with data then I'd be more than happy to update the article but if it's off by a factor of 2 or so then that's good enough for 'government work'. I'd still update the article but the conclusion would not be much different. The GPUs alone could consume that much power depending on the model but it's quite hard to find hard data on what went into the making of AlphaGo on the hardware side.
    - abecedarius 11y ago
      Geoff Hinton at http://www.macleans.ca/society/science/the-meaning-of-alphago-the-ai-program-that-beat-a-go-champ/ http://www.macleans.ca/society/science/the-meaning-of-alphag... said "I wouldn’t be surprised if it was using hundreds of kilowatts", in the same range as the other estimate I mentioned above (a couple hundred). I don't know either.

jacquesm 11y ago

$10 will buy you 100KWh, not 1MWh, you're off by a factor of 10. > why haven't we yet found a productive way to spend lots of computing power on making us better programmers? Usually the most cycles are spent on compiling and on test runs, costing only a tiny fraction of our hourly rates. The Go language (ironically) addresses some of this by putting programmer productivity very high up in their feature list by focusing on compiler speed to the point where some of the differences between compiled and interpreted languages disappear.

DonaldFisk 11y ago
Turbo Pascal also had very fast compilation times (no more than a few seconds for large programs) 30 years ago, and it worked well then. But you don't need to compile the whole program, just the functions you have changed. With Lisp incremental compilation is often used, and it is almost instantaneous - just point the mouse at the function you want to recompile in EMACS, hit the load "chord", and it's ready to run in the listener. You're not even aware the compiler is there. Of course, there's a lot more to programmer productivity than compilation speed, as Fred Brooks pointed out.
abecedarius 11y ago
Oops, thanks for the correction. OTOH zamalek elsewhere in these comments thinks 1MW is a considerable overestimate.
- jacquesm 11y ago
  The 1MW figure was based on what I could get out of the available articles, if anybody has a better estimate backed with data then I'd be more than happy to update the article but if it's off by a factor of 2 or so then that's good enough for 'government work'. I'd still update the article but the conclusion would not be much different. The GPUs alone could consume that much power depending on the model but it's quite hard to find hard data on what went into the making of AlphaGo on the hardware side.
  - abecedarius 11y ago
    Geoff Hinton at http://www.macleans.ca/society/science/the-meaning-of-alphago-the-ai-program-that-beat-a-go-champ/ http://www.macleans.ca/society/science/the-meaning-of-alphag... said "I wouldn’t be surprised if it was using hundreds of kilowatts", in the same range as the other estimate I mentioned above (a couple hundred). I don't know either.