"The Reality of TDP Ratings"

You're making several mistakes in your comparison between TDP ratings and real life power consumption measurements here. You're applying the same kind of reverse logic on the SPECpower results that I've noticed before in one of your recent blog posts, also regarding power consumption. I'd like you to have a look at these SPECpower results:

http://www.spec.org/power_ssj2008/results/res2008q2/power_ssj2008-20080506-00053.html
http://www.spec.org/power_ssj2008/results/res2008q2/power_ssj2008-20080528-00062.html

These two systems scored entirely different power consumption results despite the fact that they're both using the same CPU. The DL180 G5 draws significantly less power than the DL160 G5, even though both systems should be under equal load, considering they're achieving almost identical performance results. That big difference (44W/23%) cannot be accounted to the difference in RAM setups alone.

The SPECpower_ssj2008 results apply to entire systems with full disclosure of the test details and as with all SPEC benchmark results those details are there to be taken into consideration. And that's where you're making your mistakes: You're neglectic important aspects (wether out of ignorance or intent I do not know) and are reducing the figures to a comparison between CPU specified TDP and measured power consumption.

As for the above two systems, an important difference is that one compares a 1U and a 2U solution, in which the former clearly has a disadvantage in keeping the components cool. Look at the difference in minimum ambient temperature.

Therefor, "System designers use the TDP rating to figure out what kind of a heat sink and airflow they need to design around a CPU. It tells you nothing about how much power a CPU or system consumes in real life.", is a misleading statement and your attempt to prove it based on the SPECpower results is simply wrong.

The TDP really gives you the maximum power draw under a real life workload (not some theoretical state in which all parts of a chip are simultaneously fully active, which could never happen) and worst case thermal conditions (the cooling solution operates at its limit - that's where the thermal design comes into play). There's no single benchmark which stresses all individual components of a system simultaneously at 100% load. I don't believe that would be possible even if someone would try to design such a benchmark. SPECssj is just one kind of load and it's definitely not suitable to weigh measured results against design specifications. It's even less sensible to try to reduce the differences between two systems to the power efficiency of a single component (CPU).

"but most of that discrepancy is because Intel 45nm CPUs come nowhere close to the TDP rating, and they're far more efficient than the TDP suggests."

It's common knowledge to anyone interested in the latest PC hardware, that Intel's 45nm CPUs are very power efficient and I don't want to be mistaken as arguing against that. I don't. But your conclusions are again wrong. The TDP still relates to the maximum power consumption in the same way it does for all chips. Intel isn't mysteriously sandbagging the performance efficiency of its products. There are probably some marketing issues which statistically affect measured results (binning vs. yield), since Intel has no reason to push the limits of their manufacturing potential (overclocking tests clearly show that), but more important is a well-known technical aspect: shrinking semicondunctor features increases leakage. And while at 90nm leakage only rises linearly with temperature, at 45nm it does so quadratically. So it's that simple: For the same maximum Tjunction, Intels 45nm parts may draw as much power as their TDP specifies, but under typically much better thermal conditions, they will draw significantly less than 65nm parts with the same TDP.

So please stop telling everyone they should forget about the TDP and only believe individual tests, unless you understand the implications yourself and explain them.

If you are NOT a fan or chassis designer, stop looking at the TDP.

"So please stop telling everyone they should forget about the TDP and only believe individual tests, unless you understand the implications yourself and explain them."

If you are NOT a fan or chassis designer, stop looking at the TDP. You should completely forget about AMD's ACP because it is a joke. AMD telling people that AMD ACP is most similar to Intel TDP is one of the most blatant lies I've heard in recent years by a technology company.

CPUs can come close to the TDP under extreme workloads that are very optimized and utilize every portion of the CPU. LINPACK for example comes very close to stressing CPUs to the TDP. But in the real world, you only need to care about your particular workload which will fall far shy of TDP. In the real world, you worry about the total system power under your own worst case load.

You're making the same mistakes with measured results that one may make looking at the TDP

I didn't mention ACP and I don't know why you're bringing it up. It's of course a marketing metric, although I haven't seen AMD truly lying about it's nature or claiming it to be comparable to Intel's TDP.

The TDP is the maximum power consumption under real workloads and extreme conditions and I'm glad you finally admit that it is related to the electrical specifications of a component and not just to its thermal characteristics. As long as one is aware of all the implications of this metric and doesn't have more accurate results for his particular usage scenario, it therefor remains a good guideline for power consumption. Not for calculating absolute values, but for estimating the top margin of worst case results.

You're on the other hand making the same mistakes evaluating measured results that one may make when applying TDP values. SPECpower_ssj2008 provides you with results for entire systems under one specific workload and a single set of (not necessarily comparable) conditions. Read the methodology PDF provided by SPEC to learn how to interpret the results. This is in no way suitable to reduce the problem to individual components or disqualify design specifications as useful guidelines. I've given you links to two results for the exact same CPU model that differ materially in power consumption.

Evaluating power consumption is difficult and there are no simple answers, like "systems using this CPU will be more power efficient". SPECpower_ssj2008 is a step in the right direction and is very useful for people running similar workloads on comparable hardware. It doesn't help at all when you're planning to equip your office space with client machines. We have yet to see a benchmark that accurately measures power consumption for office workloads on desktop machines. Most hardware enthusiast sites test under the wrong conditions and only provide results for idle systems and something which they consider full load. I haven't seen a single test properly reflecting the effects of modern equipment's power saving features, which only come into play at partial workloads (AMD K10's and Intel Nehalem's individual clock domains, for example). Until there are better benchmarks, the TDP, a proper look at a certain product's characteristics and the careful examination of several more or less accurate tests is the only way to get a good understanding of what might be the "greenest" solution for a particular task.

Re: How to achieve Green IT

actually the only issue I have with that article is that you don't mention the power consumption of the air-conditioning units in the business as well. All that power wasted as heat has to be extracted from the building in the air-con and that equals yet more electrical power. I understand why you didn't mention it (as IT management almost never sees power bills) but it's one thing that I always force myself to consider when recommending hardware updates and in particular the merits of virtualizing servers, or with a thin-client office environment.