On Jan 5, 2016, at 3:42 PM, Valeri Galtsev <galtsev at kicp.uchicago.edu> wrote: > > And yes, there is downside in keeping older hardware around: wasting > precious server room space, power, AC. Out of interest, I calculated a common case. My recollection is that 10 years ago, a typical entry-level headless server would draw about 100 watts at idle. I measured an entry level server a few months ago: it drew only about 35 watts at idle. Power costs about 14 cents per kWh here, so 5 years of 24/7 use comes to about US $400 in electricity alone. Double that during summer to account for air conditioning to remove the heat generated, with less power needed in other parts of the year, and we’re talking about more like $600-750 in extra power costs to run that server for 10 years instead of 5. Watts/MIPS continues to drop exponentially: http://www.singularity.com/charts/page129.html Whether your costs also drop exponentially depends on the growth of your MIPS used per year. That is, you get all the benefits implied by the graph only if your MIPS used remains constant. If your computing use is outstripping Moore’s Law gains — i.e. your processing centers are getting physically bigger or drawing more power than ever — the benefits merely offset or enable your growth costs. If you’re like us, however, in that new computer power isn’t instantly sucked up by increased load, so that newer computers still feel faster than older ones, the better sleep mechanisms in OSes and hardware mean that you get back to an idle state faster than before, so some of the benefit from faster hardware shows up as further increased power savings. If that seems implausible to you, run the curves to their [il]logical extreme: an infinitely-fast CPU that draws 0 W at idle but spikes arbitrarily high when processing computes your problem instantaneously, so appears to draws 0 W continually. And Xeno never reaches the wall. ;)