"And in what way did you allow for the proven fact that Intel's TDP isn't the maximum power-output for their processors either? It's been proven that an Intel Processor can draw and dissipate more power than their TDP lists.

If it weren't for this discrepancy which has become a major marketing hurdle for AMD each time it is encountered there would have never been a reason to introduce ACP. In short, you're barking at the wrong tree for the cause.

I concur with you that their averaging method is questionable. But barking solely at AMD for the existence of power-requirement number that don't easily translate to a maximum of let you calculate your own average you should take the blame and pass it on to Intel."

F Masolijn

"When it comes to power, the only approriate mean is the arithmetic mean, which accurately tracks what the meter does. Introducing the geometric mean is pointless.

It is essential to guestimate the amount of time spent in each mode, and that is very variable depending on the payload. For business and web server payloads, most of the time can be spent idle, especially if the CPU and OS cooperate nicely to this effect (a single poorly written tiny app can change that drastically).

If there is any scientific workload around, nearly all of the time is spent near peak power: scientist are power-hungry."

Francois Grieu

"The Average power issue is a total crock. The entire point behind multi-core servers is to allow virtualization to consolidate onto fewer systems. AMD touts this as a benfit of the Barcelona - and that's not a lie.

The problem is in a virtualization utopia every system would be utilized 100%. Schemes like this are coming with VMMotion to move virtualized machines around to use free resources and even split a VM across physical boxes to leveral spare unused compute cycles.

So, if this virtulization panacea is the playground of Barcelona, then how does average power consumption play? If a machine is 100% utlized, as is the goal, then average power dissipation = maximum power dissipation...

AMD is playing us for the fools..."

Deepthroat

Finally, a man with a fine understand of such technology puts fingers to keyboard and shows you all how it's done.

"Your article on AMD's claims on Barcelona Power seemed to me to contain less real information and more inuendo than the original slides must have... although my understanding is hampered by not having seen the slide deck ;-)

I can understand your complaints about their use of the geometric mean rather than the arithmetic mean (more on that later). But in the rest, it appears to me that you have missed the point rather than they. I know averages are easier to manipulate than TDP, but they are far more useful to everything I as a consumer need to know. In particular, it is some kind of average which tells me how long my battery will last and some kind of average which tells me how green is my system. These are the two questions I want answered about power, as a consumer. TDP means nothing to me except what shows in the price tag anyway, so I don't care, as a consumer. (In my day job as a processor designer, of course I care about TDP because of its effects on packaging cost, reliability etc. but those are subjects unrelated to the consumer's concerns about power.)

As to how to answer those two questions, my battery life depends on the average inverse power (seconds/joule averaged over my workload). This is commonly known as the harmonic mean. How green my system is might best be described by average power. Average in the sense you used it (aka arithmetic mean). Now, as AMD, I could quote both of these numbers, but that would really confuse almost everyone. And it's not always clear which one wants for which purpose, either.

AMD have chosen to quote the geometric mean, which has two useful properties. First, it is usually between the average (or arithmetic mean) and the harmonic mean and isn't usually terribly far from either (although it's possible to make it far). Second, geometric mean has the property that for say a mean of 10 numbers, a 10% change in any given number is the same as a 1% change in all numbers. That is, benchmarks are equally weighted without choosing weightings and remain equally weighted after a change. With any of the other versions of mean, changing the score on one benchmark changes its representation in the mean.

So I would argue that, if AMD wish to address the most important issues, they have chosen the best single number with which to address these issues. I think you would have been much better off spending your time showing where they have manipulated the results or used benchmarks which are more favorable to AMD than what most customers run rather than generally complaining about something for which their presentation (as best I understand it through your article) makes far more sense than you do."

Bill Huffman

So there you have it - at best, AMD's new benchmark metric is confusing. At worst... well, we daren't speculate for fear of igniting the Intel fanboy claims again. Stay tuned to watch this one run and run. µ