The Inquirer-Home

Average power consumption is anything but

Smoke and mirrors lie behind Barcelona's numbers
Mon Sep 10 2007, 23:26
TODAY IS the day, folks. Today is the day that, after all the months of speculation, AMD's Barcelona chip finally gets its official unveiling at the hands of Mr Ruiz.

Of course, AMD has been briefing on the capabilities of these new chips for weeks now, and has certainly been working hard to make sure that its 'corporate comms' are 'on message'. With that in mind, we grabbed AMD's pre-brief slide deck, the PowerPoint work that the company has been using to illustrate the performance of its new platform. Reading it is only slightly preferable to doing shots of washing up liquid, but you know us, here to do the hard work so you don't have to.

We were hoping that the interwibble would have stumped up plenty of benchmarks and independent analysis of AMD's technology by now. Alas; this has not been the case. It seems that, bucking the trend most hardware manufacturers tend to follow, the folks from Austin have declined to put early systems and silicon in the hands of testers, instead using their slide deck to tout performance and feature advantages found in their own labs. A dubious practice to be sure, such conservatism warrants a closer look.

For the most part, AMD's briefings have focused on the advantage Barcelona has in terms of power consumption and featureset over Intel's quad-core Xeons. Much has been made of the fact that this is the first 'native' quad core part, without any real discussion of whether this is a bona fide benefit. Yes, it is in theory a 'cleaner' architecture. But an architecture lives or dies by its performance, so are theoretical facets really worth a dime?

The key factor that AMD is placing its bets on is its low power consumption. Well, in a bid to change the way we think about power, AMD has unveiled a new metric it is hoping we analysts will buy into - average power consumption. This is not, in itself, a bad idea - it attempts to convey what the 'average' consumption of a platform is, rather than relying on the worst-case-scenario TDP, thus enabling system builders to get a more accurate representation of system performance.

Of course, 'average' is a loaded word. What does it actually represent? Delving into the testing methodology, scantily given at the end of the deck, we discover that the average, in this case, is the geometric mean of the amount of power the processor is drawing from the rails. Is this misleading? The geometric mean is different from the arithmetic mean, which itself is different from the median or mode, as any high school maths student with tell you. If we take an average workload of six minutes, assume that one minute is spent with the processor drawing each of the following powers:

80, 80, 80, 105, 120, 100W

The arithmetic mean - adding them all up and dividing by six - gives you an average power consumption of 94W. The geometric mean - finding the product of them all then the sixth root - gives you 92.95W, a lower number (the geometric mean, which AMD uses, tends to skew lower, conveniently).

Of course, both of these numbers are totally useless. AMD consequently advises that designers should build their systems around a 92W workload - even though the processor, during this six minute period, is never at 92W. The average workload under these measurements is a purely theoretical number that bears no resemblance to the real world performance of the chip. Either designers should design for the most common workload - here, 80W - or the maximum workload - here, 120W. Inventing a performance number that doesn't exist is hardly a great real-world performance benchmark.

Indeed, the only real power number that system engineers care about is the amount of power being consumed at the wallplug over a 24 hour period, a number which can be usefully averaged, unlike this processor power claim. In failing to grasp this, AMD succumbs, unfortunately, to the allure of creating another easy-use number that means nothing, not unlike a 3DMark or SPEC rating - easy to tout, utterly useless in the real world.

And there are plenty more of these to be found in the rest of the AMD deck, the firm consistently relying on SPEC results and test HPC loads to drum up the necessary Intel-besting benchmarks. Of course, very few of these reflect any kind of real-world scenario, and the majority of the tests rely on incredibly high-end, processor intensive workloads - workloads which are unlikely to be replicated in the real world amongst your average systems administrators.

Of course, we wouldn't expect any product manufacturer to put out benchmarks showing its products in anything but their best light - but in an age where websites with independent analysis are ten a penny, the exercise seems at best futile, and at worst misleading.

This is something Intel learned a few product cycles ago, and which AMD needs to learn fast. Dare we suggest that the reason we are lacking independent analysis of the new Barcelona chips is because so few have been sent out to the press in a bid to see some semblance of a meaningful quantity shipping to customers?

Indeed, with no products to get to grips with, the fanfare for AMD's hope and saviour chip has been, well, muted to say the least. A look around the wibble this evening sees most top-tier IT sites not even leading with the Barcelona story. We would say it is too early to call the Barcelona launch a squib, were it not for the fact that AMD touts Barcelona's 'rapidly ramped' development and 'successfully delivered' roll-out.

Really? µ


Share this:

blog comments powered by Disqus
Subscribe to INQ newsletters

Sign up for INQbot – a weekly roundup of the best from the INQ

INQ Poll

Heartbleed bug discovered in OpenSSL

Have you reacted to Heartbleed?