The popular new PPW (performance per watt) measure of value for CPUs is snake oil, no matter which vendor's brochures and billboards hawk it. At least IT has the convenience of a two-variable formula to determine where your hardware dollars should be spent.
Experience? Judgment? Bah. Who needs these when all you really need is ((clock speed) x (instructions per cycle)) / watts?
AMD did itself a mighty disservice by buying into Intel's ridiculously simplistic measure of total system value, and by doing Intel one dumber by applying a formula originated as a consumer come-on to the enterprise. Now Intel is turning the tables on AMD by saying that Intel will focus on multicomponent PPW, not just the CPU's.
That's interesting because, when Intel's whole PPW campaign started heating up, people I spoke with at AMD were incensed that Intel was making the CPU the sole focus of the PPW formula. They argued that Intel's systems are no great champions of efficiency when measured as a whole. But when AMD bought into PPW, it applied it to the CPU alone. Intel, however, has already moved on. It makes me want to ask AMD, "Didn't you guys see this coming?"
PPW puts a clown's wig on a serious issue. IT has to strike the right balance among performance, cooling demand, availability, power consumption, and noise generation. That takes study and experience, not first-grade math. PPW gives less-experienced IT planners the impression that the job of pulling together an efficient server room is simpler than it is.
It isn't simple. There is no formula for it. What there is instead is a thorough study of the factors relevant to your operation. For example, if your approach to dealing with demand-spikes and potential failures is to run hot fail-over servers, it's time to look at new approaches so your systems aren't waiting around for a crisis. Do you have infrastructure peripherals that you know won't be used during certain hours or that can be brought online from a suspended state when needed and suspended again when they're idle? Do you have multiprocessor and multicore systems that have all of their cores active even when the system has little or no work to do? These are the sorts of capabilities that make me such a fan of system, storage, and infrastructure virtualization. I've got a shell script running on my Xserve G5 that turns off the second CPU until load rises above a set threshold. Likewise, my Xserve RAID suspends itself if it's been unused for more than two hours. These sorts of things are not hard to do if the vendor makes them possible.
And that, right there, is the key. You've got the experience to know that cooling, power, and noise are vitally important, and you're the only one who can gauge your need for computing, storage, and networking infrastructure. IT doesn't need formulas; let AMD and Intel wrestle that out in consumer systems.
IT doesn't need to get sucked back into the useless gigahertz wars. IT needs vendors to supply worst-case, raw facts that help it judge the efficiency of systems that it already knows will do the work at hand. And then IT needs assurance from vendors that planners and administrators will have the capability to control the power characteristics of deployed systems at a finer level of granularity than on or off. But that's a subject for another column.