Originally Posted by factotum
According to another article I found, this machine uses 9MW of power at full load. At the sort of prices they charge for electricity in the UK (approx 20p per kWh), that would be £1800 an hour or a little under £16 million per annum. Now, what I *can't* find is how much this machine cost to build in the first place--I suspect that number is going to be very high compared to that estimated electricity cost, though.
That's significant, because it goes back to what I was saying earlier--if the machine spends most of its time idle you could have built a less powerful, cheaper one to do the same task.
I'm going to define "idle" as "0-1% utilisation" at this point, just so we're all on the same page. Correct me if your definition if idle is different, and we can revise from there.
Anyways, back to cost-benefit analysis. Some rough Googling presents a $1.2bn
price tag for Japan's K supercomputer, currently #2 on the TOP500 list. Peak consumption is 12.6 MW, though further down the page, it lists 9.89 MW, so let's go with the latter number as the expected figure.
They're comparable I suppose. I'd have to do some more research on Titan though.
I do have to say though, that the arguments presented thus far comparing supercomputers and regular consumer computers don't really apply. First off, any large computing center will probably be running around 60-70% load, but that's an aside (remember electricity scales exponentially). The average user might think "oh, a bigger computer means a faster computer!" when that really isn't always the case. Sure you can use a big computer to accelerate your programs, but what's of more interest is that a bigger computer allows you to solve bigger problems
. Some programs only gain a limited benefit from tacking on more cores, and need an alternative approach when switching to parallel.
It's not about building a less powerful, cheaper one to do the same task. Sometimes it's about building a bigger machine that can handle a larger-sized problem, courtesy of the terabyte or more of RAM attached to the thing. Atmospheric simulations come to mind here, if you're trying to model a very large region (or the entire planet, in Climate simulations).