IBM Announces $1 Billion Linux Investment for Power Systems

Posted Sep 22, 2013 18:45 UTC (Sun) by giraffedata (guest, #1954)
In reply to: IBM Announces $1 Billion Linux Investment for Power Systems by khim
Parent article: IBM Announces $1 Billion Linux Investment for Power Systems

How about checking the facts instead? Quite expensive supercomputers are often decommissioned because of huge power bill.

That particular fact is irrelevant to the question of whether it's more important for a computer with a million processors to be energy efficient than one with one processor. Small computers are also decommissioned because they are energy inefficient. Big supercomputers are also kept in service in spite of the fact that they have huge power bills.

The large supercomputer in your example is being decommissioned because the computations it does are worth less than the energy it takes to make them. I have shut down single-processor machines in my house for the same reason.

Situation is radically different WRT supercomputer then WRT your desktop computer not because it's million CPUs vs one but because it's million loaded CPUs vs one unloaded CPU.

So in spite of starting off by saying my statement that the large number of CPUs isn't what makes energy efficiency especially important in BlueGene/Q, here you agree with me by saying it is instead the high utilization of CPUs that makes energy efficiency especially important in BlueGene/Q.

But I can't see that either, except insofar as the cost of energy efficiency, or of energy, doesn't scale with utilization (my point).

If your computer cost $100 million and its power bill is $9 million then in 5-6 years you'll pay for the power half of what you paid for the computer itself!

Is there reason to believe it's a different ratio for a machine with fewer processors?

There is nothing to divide: it's one power station, it supplies you with power, you pay for it, end of story.

Again, you're talking about unscalable energy costs, and thereby supporting my claim.

You're assuming the supercomputer will be powered by spare capacity in an existing power station. That means energy cost doesn't scale with the number of processors: Doubling the number of processors might increase the energy cost by a factor of 10 (you have to build a new power station), so it makes sense to make more expensive, more energy efficient processors.

That probably isn't realistic, though. Power stations are being built all the time, especially when there are people around who want to plug in supercomputers. In the big picture, assuming energy supply is inelastic probably doesn't lead to the right technology.