AMD and Intel power FUD doesn't add up Both are right, both are wrongBy Charlie Demerjian: Friday 23 June 2006, 10:14 THE FUD IS swirling hard in the Performance Per Watt (PPW) realm. AMD and Intel are getting into a piddling match of epic proportions. Both claim wins, and both claim the other side is being flat-out disingenuous with its claims. The problem is they are both arguing completely different things, arriving at a number that only has the units in common with each other.Lets look at what they are claiming. AMD says if you add up the TDP of all the parts of the system, AMD is a lot better than any Netburst CPU. They also claim leadership over Woodcrest systems with the 'add the TDPs' measurement.To make matters more confusing, both companies measure TDP in a different way, oh joy. AMD measures it as the maximum power that a CPU can possibly draw while Intel calls TDP the maximum you are likely to see sustained. One Intel TDP ~= .85 AMD TDP. To correct for this, AMD bumps up the Intel numbers to max draw.Intel says this is complete bunk, you can't get to the maximum draw on all components, and even if it was possible, it is unlikely to happen in the real world. Instead, they say you should measure average power, and when you do, there is a clear win for Intel and Woodcrest.Both arguments are correct in several ways, and more importantly, both are seemingly purposely overlooking some key facts. In fact, I would go so far as to say both are wrong to the point of near uselessness.Lets start out with the fact that the Woodcrest chips are not out yet. OK, they are out next Monday, but you can't buy it now if you wanted to. A few weeks after the Woodcrest chips are out, the 1207 AMD parts are coming out with DDR2, significantly dropping power again, so any numbers before August 1 are again flawed. Basically, until I get a production 1207 rig and a production Woodcrest rig in here, I don't believe any of the numbers.The next problem is that Intel is right, you can't make all the parts pull their TDP, or at least AMD TDP. If you are having the CPU churn at full tilt, you are probably not idling waiting on memory, so that is underutilized. If you are waiting on HD accesses, an eternity in CPU cycles, again, you are not banging the CPU hard. Fair accusation.To counter this, you can say that it is unlikely, but if you are designing a data center, you have to design for a major portion of that number, usually a high fixed percent of maximum TDP. TDP stands for Thermal Design Point, and you need to supply that much electricity and remove that much heat or bad things happen. If you cook 100 racks of equipment on a hot day because the data center was under-specced, I don't think your boss will buy the 'It was unlikely' excuse. Moral here, unlikely but you have to take it into account.The Intel argument of average use is also bunk because what is average use? Is average an HPC machine running a load hand tuned for the box? Is it a gamer banging away at BF2? I person sitting at the office typing on Open Office, or maybe watching a movie? How about both? What is average, and how can I test it repeatably? Answer, you can't.Then there are the measurements themselves, somewhat flawed all told. You can't buy either system that matters yet, but that aside, there are still problems. AMD bumping up Intel power numbers is pretty bad. Even if they did it fairly, you can't fiddle with the other side's numbers because they don't fit. In this case, AMD should have re-measured their numbers using Intel's TDP scheme instead. As a semi-aside, AMD still will not publish TDPs for individual chips, an astoundingly annoying habit.AMD also guesstimates the power for portions of the north bridge it includes on the chip but Intel has externally. The numbers I have seen for this seem pretty good, but in a system measurement, it is irrelevant. Not the end of the world, but an opening for criticism.Intel on the other hand completely ignores Netburst and concentrates on the new. The systems are specced out to the bare minimum, they use four FBDs, and AMD has things configured to the max. One way favors Intel, the other favors AMD. In a shock to no one, each sides pick the most favorable. Curiously, Intel picks a number of FBDs, four, that does not give peak performance, that is at eight. Those eight brings with it a power penalty of about 50W, enough to put it over AMD in the ratings. At 16, AMD cruises to victory like they claim.So, you have both sides arguing a completely different set of numbers, with completely different configurations, and coming up with a number that they use to directly compare one to the other. As my grandmother would say, paraphrased, those disingenuous shorthand for FileSystem ChecKs.The papers put out by both sides are again pretty flawed for the same reasons I listed above and one more. The AMD ones compare TDPs, useful if you are designing air conditioning for a data center, but less so for the real world. Intel quotes numbers that are unrepeatable, and probably don't match what you are doing. Some of the AMD slides contain a handful of estimations, and usually ignore Woodcrest.One paper Intel is touting now* has Powernow turned off because of problems running it under Windows64, and to be fair, they turned off SpeedStep. Given the circumstances, the conclusions under load are solid, but may change if the power savings are turned on. The idle numbers most certainly will. Either way, I would hold off on extrapolating from the data until AMD clearly gets the word out about PowerNow on Win64.Either way, what it comes down to is this, until AMD 1207 parts are out in the wild AND you have tested both with the apps you use in the configs you use, ignore both sets of FUD. Neither set fully applies to the real world, and both are so skewed they are borderline laughable. µ(*) System Power Unequal To Sum of The Parts, June 22, 2006, Thomas Weisel Partners
