This is a question that has been on my mind a lot recently. It applies to CPU figures as well as GPU figures. I think smilingcrow has it right: They're not terribly reliable, and they are certainly not representative of typical usage.

I think this issue has already been dealt with, at least where cooling is concerned. SPCR has known for years that CPUBurn is one of the heaviest loads that you can use to stress a CPU; realistic, non-benchmark loads do not eat up as much power or generate as much heat. The purpose of using CPUBurn is to test heatsinks in a worst-case scenario. If a heatsink can handle a processor running CPUBurn, it will be fine for other strenuous activities. In addition, it is such a simple load that it makes testing easily repeatable — not something you can say about "typical" usage scenarios.

Knowing the maximum power draw has a similar use in the context of GPUs — it helps single out worst-case scenarios for testing. It was Xbitlabs' testing that helped me recognize that the stress-test that we were using (ATI Tool) wasn't stressing the X1900XTX enough to trust it as a benchmark.

However, when it comes to comparing different GPUs (or CPUs for that matter) with each other, I don't know that I can think of any feasible test method that I would trust. Using the same utility to test all processors (as we have been doing for some time) is less than ideal because different chips have different characteristics — the most stressful load is different for every chip. If we're serious about comparing a maximum power figure (vs. a typical figure), we need to use a customized utility for every processor.

Let me talk about CPUs for a bit, since it will allow me to use CPUBurn as an example. One of the nice things about CPUBurn was that it includes different utilities for P6 (P-III?) vs. K6 architectures. The fact that Intel and AMD's later chips were built on these earlier architectures has helped CPUBurn remain a useful tool, out of date as it is. Subsequent testing of the utility, which showed that no greater load could be found, confirmed the usefulness of CPUBurn.

However, with the release of Conroe, which is a significantly different architecture, things have changed. 2xCPUBurn no longer even approaches the heaviest load it is possible to put on the Conroe. Our testing showed it to dissipate about 35W — unbelievable on a chip with a TDP of 65W. To hit 65W, I would guess that we need a utility that simultaneously loads all of the chips SSE units, plus occupies all 4MB of cache. As far as I know, this utility does not exist. Real world encoding applications may come close, but ultimately they're not easily repeatable and are unlikely to be optimized to use 4MB of cache.

So, not only do we not have a utility that we can trust to give us maximum power dissipation on Conroe, but the usefulness of such a utility is less than it has ever been. Why should maximum load matter now when
a) typical usage is so far below, much lower than it has been with Intel's previous processors if CPUBurn is anything to judge by, and
b) the heatsink market is filled with products designed to cool twice the TDP of the Conroe.

It seems to me that the maximum power approach is really only useful when judging what is needed to cool a processor, and, given the two points I just made above, it is less relevant now than ever before. When comparing heatsinks, it is really only necessary to maintain the same chip/load combination for all heatsinks tested; maximum power is not a necessity in that case.

If your interest is power itself (i.e. energy efficiency, not cooling), then maximum power is of little interest, since no modern chips run at full load all the time (I have a GeForce 3 that seemed to ... but I haven't seen any more recent examples).

For efficiency, it is far more useful to come up with a ballpark "typical" figure; for most people this is probably very close to the power consumption at idle. However, trying to test for typical usage is an exercise in futility. Nobody actually has "typical" usage (no family has 2.3 kids), and discovering and trying to characterize how different usage patterns vary from "typical" would require a wide-scale study. Even if such a study were done, the practical use of the information seems very low to me — making use of the information would require a close study of one's personal computing habits... and how many people do you know who could/would do that?

It seems to me that GPUs are very much in the same position as CPUs. nVidia's and ATI's chips have different power profiles, and, even if a steady state "maximum power" load could be found for each of their various processors, gaming is such a dynamic activity that such a load would be useful only for testing coolers, not judging differences in power consumption.

...

This is a question that has been on my mind a lot recently. It applies to CPU figures as well as GPU figures. I think smilingcrow has it right: They're not terribly reliable, and they are certainly not representative of typical usage.

However, with the release of Conroe, which is a significantly different architecture, things have changed. 2xCPUBurn no longer even approaches the heaviest load it is possible to put on the Conroe. Our testing showed it to dissipate about 35W — unbelievable on a chip with a TDP of 65W. To hit 65W, I would guess that we need a utility that simultaneously loads all of the chips SSE units, plus occupies all 4MB of cache. As far as I know, this utility does not exist. Real world encoding applications may come close, but ultimately they're not easily repeatable and are unlikely to be optimized to use 4MB of cache.

SPCR has known for years that CPUBurn is one of the heaviest loads that you can use to stress a CPU; realistic, non-benchmark loads do not eat up as much power or generate as much heat. The purpose of using CPUBurn is to test heatsinks in a worst-case scenario.

About maxing out a C2, would it use more power if you run CPUBurn, 2x P95 (not set to highest priority), and some other apps at the same time? Or would some kind of priority still make the same result (read power consumption) as a single app run?.

The CPUTest from 3DMark06 might be a good candidate, but it is also nice if the test like Prime95 will actually test for errors--as merely a system merely continuing to run doesn't mean it is running properly. Some test for thermal throttling would also be good--Netbursts were certainly capable of merely throttling themselves down.

I think you guys are missing the thrust of Devon's post. The point isn't "how to get the maximum heat out of a CPU for testing" it is, "Why are we trying to get the maximum heat out of CPU for testing?" If a CPU or GPU can never run as hot as during the testing, what's the point of testing it under those conditions?

Even if you find a better solution for general CPU testing purposes, utilities such as Prime95 which use error checking are still very useful and I would say necessary for system testing. This is true whether over-clocking or under-volting. Since SPCR does under-volt in certain of its tests, you’ll still need to use such utilities at times.

Utilities like that are useful if you are testing a CPU, but CPU errors are meaningless if you are testing a Heatsink. For a heatsink the only things that matter are °C/W and dBa. Now if it was a CPU review where you were testing overclocking, underclocking, unvolting abilites, etc, then error checking would be important part of the methodology. But to a heatsink, the only thing that matters is it's input wattage. Think about it this way...if in our testing that our CPU had Prime95 errors while testing with a Brand X heatsink. Is that result meaningful? No, because CPU's are individually unique...just because our had errors doesn't mean your's will, and the heatsink probbaly had nothing to do with it.

The idea of uses "real" apps to get typical wattage information seems like the right track, but even that is full of pitfalls. Those apps will typically run as fast as the CPU will let them, so it doesn't really help us come up with a single "typical" wattage to use for testing. Do you test with a wattage that equals the fastest CPU out there, or the median? I think an easier way would be to pick one arbitrary wattage to test heatsinks at, and then allow the community to compile a "typical wattage" of actual CPU's. Then a user could look at what their individual wattage is in comparison to the "testing" wattage and extrapolate what the temps would be in their system if they bought that heatsink.

GPU testing is equivalent to CPU testing, really. You just swap "app" for "game". I think you're on the right track of using gameplay to come up with a wattage. Do the testing with a few different cards and a pattern may emerge, something to the effect of, "Doom3 averages a GPU wattage that is 75% of the GPU's theoretical max. 3DMark is 95%, etc" That sort of testing you'd only have to do once, not everytime you testes a new heatsink.

Can of worms..can of worms indeed.