Debunking Power Supply Myths
by Christoph Katzer on September 22, 2008 3:00 AM EST- Posted in
- Cases/Cooling/PSUs
Connectors and the 12V Rail Issue, Cont'd
As a second example, let's take our high-end system. This requires more power, making it a little more difficult to find an appropriate power supply -- especially if you want to run 3-way SLI. NVIDIA initially only certified 1200W power supplies for that sort of setup, but Corsair was eventually able to receive this certification for a 1000W unit. To run a 3-way SLI setup, we need at least six 6-pin PEG connectors, so let's start by looking at the number of connectors available on our high-end PSUs.
| PSU PEG Connectors | |||
| Manufacturer and Name | # of 6-pin PEG Connectors | # of 6/8-pin PEG Connectors | Total |
| Antec NeoPower Blue | 1 | 1 | 2 |
| Antec Signature | 2 | 2 | 4 |
| Silverstone Decathlon | 4 | 0 | 4 |
| Zalman ZM850HP | 2 | 2 | 4 |
| Enermax Pro82+ | 0 | 4 | 4 |
As you can see, none of the power supplies we listed are able to run more than two GPUs, since none of them have the available six PEG connectors. We still don't recommend 3-way SLI, has the advantages over two-way SLI are sketchy at best. Regular SLI also has the advantage of only requiring four 6-pin PEG connectors even for the highest-end GPUs. All of the listed power supplies can meet this requirement, except for the Antec NeoPower Blue.
For the second 12V rail issue, here's a look at the specific power requirements of our graphics card, the GeForce 8800 Ultra:
| GPU Power Requirements by Connector | |||||
| Vendor and Chip | Through 6-pin Jack |
Through PCI-E Slot |
Total Power | ||
| NVIDIA GeForce 8800 Ultra | 9.4A | 112.8W | 6.1A | 72.2W | 186W |
If you plan to run multiple cards, you can just multiply 186W by the number of cards in the system. We found during testing that the cards don't consume equal amounts of power. The first GPU always uses the same amount of power, whether in a single or SLI configuration. The second and third cards however require much less power, which also means that they are doing less work. In testing a variety of benchmarks and games, we saw that the second and third graphics cards only use 40% to 50% of the power of a single GPU.
One thing is obvious when looking at the power consumption of the single GeForce 8800 Ultra: the maximum power draw is only 15.5A, which you can get from a single 12V rail. Most higher-end power supplies have multiple 12V rails, which should make distributing power to your GPUs even easier. As long as the 24-pin ATX connector and the different 6-pin PEG connectors run on different 12V rails, none of the rails should end up with an excessive load. That takes care of the 300W power requirements mentioned by GPU vendors.
If you take the high-end system (with an optional third graphics card), you are looking at one of the most demanding systems available. Of course, you could always overclock the processor and graphics cards, which might increase power requirements by another 30%, but we'll leave that topic for another day. We were able to run this test system with a normal 850W power supply, and even with a reasonable load the PSU fan didn't make that much noise. Most of the high-end PSUs we've chosen for this article could also run a triple-SLI setup, provided they have sufficient connectors.
At the beginning of the year, we requested a special AnandTech Edition of PC Power & Cooling's Turbo Cool 860W power supply with six PEG connectors, specifically for running a triple-SLI setup. We have been using the power supply since then with three GeForce 8800 Ultra cards, an Intel Core 2 Extreme QX6850 CPU, and 12 Western Digital hard drives without any problems. Clearly, quality of construction and number of connectors are far more important than a silly wattage rating.
Facebook
Twitter
RSS
98 Comments
View All Comments
nilepez - Thursday, September 25, 2008 - link
I've seen this site and countless others constantly promoting the benefit of 750 watt+ PSUs.The reality is virtually nobody needs 500W, much less 750 or 1KW.
AnAverageJoe - Wednesday, September 24, 2008 - link
Thank you for this article, it is really one of the more organized efforts at getting this kind of info out there that I have read and hopefully will save folks from overdoing it (saving $$ in the process). Couple of comments:1. p.1 “If people really took the time to examine system power requirements”: From the perspective of building a new rig, where does one find this information? Any sort of figures, short of the very few articles such as this and public power supply guesstimators, appear non-existent. As stated in the article something is better than nothing but getting precise information appears impossible for any given component. Both Intel and AMD publicly provide technical data on their parts but the only single figure one can really get out of that data is the TDP for the processors. I have failed to find any similar technical data for GPUs and although most graphics card reviews now include power draw it is for the system and not the graphics card itself thus there is is no way to get the discrete graphics card power draw. Ditto for every other component. In sum, I don't see any way for one to gather the required data to compute the power requirements for any planned build short of actually building the thing and putting test equipment to it.
2. I was left wanting a statement to the effect of "one important goal of choosing a power supply is to maximize efficiency under expected operating loads." The examples do make the point just took a bit longer for me to get it thru my dense skull.
Again, thanks for spending the time to put this together because this really is I think one of the most overlooked areas of system integration and in my experience impossible to more than generally guesstimate.
Cincybeck - Tuesday, September 30, 2008 - link
Try searching "power supply calculator" in Google. The one, eXtreme has, provides a pretty complete list and quite a bit of options. I punched in the low end and mid range computers from the article. The calculator is high according to the article's results, but close enough to give you a good idea.Johnniewalker - Wednesday, September 24, 2008 - link
I have been waiting for an article like this for years, thanks! I always suspected people were buying PSU's that were too big.rvikul - Tuesday, September 23, 2008 - link
Anandtech is back! This is exactly the kind of information Anandtech excels at providing. very relevant and very useful. Thanks and keep it up (and please dont go back to discussing social issues etc).OddJensen - Wednesday, September 24, 2008 - link
Strange figure for the 2900XT. Knowing it's one of the most powerguzzling single GPU cards, it should have been much higher, right?CSMR - Tuesday, September 23, 2008 - link
I'm going to repeat myself because I feel this is important.The tables claim to represent actual power consumption of processors and chipsets but the figures are very exaggerated. (There are even people who run whole systems on one or two of the chipsets listed on less power than the power the article claims for just the chipset.)
Here are more accurate CPU measurements from xbitlabs:
http://www.xbitlabs.com/articles/cpu/display/intel...">http://www.xbitlabs.com/articles/cpu/display/intel...
Behardware gets similar results. (Can't find the link straight away.)
Must fix this as Anand needs to keep its reputation for good information. You can't have figures (CPU idle power) that are out by a factor of 10!
The graphics idle power data are better than the CPU data but also too high, by a factor of about 2, compared to existing measurements:
http://archive.atomicmpc.com.au/forums.asp?s=2&...">http://archive.atomicmpc.com.au/forums.asp?s=2&...
JarredWalton - Tuesday, September 23, 2008 - link
The question is: did Christoph measure power incorrectly, or did someone else? From my understanding, he's measuring the current on the various wires leading from the PSU to the components. HDD is of course easy to measure. The ATX12V/EPS12V connector supplies the CPU, so that's simple as well. PCI-E gets 12V from the PEG connector along with the extra four pins on the 24-pin ATX connector. The remaining pins on the 24-pin ATX feed the chipset, RAM, and other motherboard components. Sum all of that together and you get the power draw for the entire system.Perhaps the CPU power draw numbers are high and the chipset/mobo numbers are low, but worst case the point of the article is to show that higher wattage PSUs are not required for most systems. A midrange system with similar components might use a bit less than what we estimate, but I'm quite sure it wouldn't need more power than our high estimate.
I won't guarantee that the individual numbers are 100% accurate, but I doubt that desktop C2D processors are idle at only 7W or less. I know on my own C2Q Q6600 8800GT 4GB system it idles at 176W power draw, and if I put a load on just the CPU (Folding@Home SMP) the draw increases to 262W. Guessing at 80% efficiency, the components are consuming 141W to 210W, which means CPU (and mobo, chipset, and RAM) power use went up 69W. That's pretty close to the QX6850 result (fudging on the 65nm vs. 45nm and Penryn vs. Kentsfield). The Fur benchmark also gives ~262W average (one core is 100% load on the CPU), so the GPU + one CPU core increases power consumption by the same 69W, but some of it goes to the CPU and most of it goes to the GPU. That jives with the 51W increase in power Christoph measured on the 8800 GT. Running both - a "worst case" test - gives a power draw of 315W to 337W, with an average of around 324W.
So a system somewhat similar to Christoph's "midrange" setup (and estimating efficiency) uses 141W idle and 260W load. That doesn't include trying to tax the HDD or DVD, which might increase the load by another 20W, and it uses a single GPU instead of a 3870X2. What's the specific amount used by the CPU, the GPU, the RAM, the chipset, and the other motherboard components? I can only guesstimate, but stepping back to examine the whole picture I don't have any serious problems with the tables on page 1. CPU power is probably lower, since it sounds like Christoph measured the current going through the ATX12V lines and some of that will feed the VRMs and other bits and pieces on the motherboard.
Christoph Katzer - Wednesday, September 24, 2008 - link
Thanks for the explanation. The point being is that I measured how much current is going to the CPU through the 12V rails. What happens later doesn't matter since the power supply needs to deliver X current at that specific time.That doesn't mean however that the CPU is actually using all of this delivered power since we loose power at the VRMs and CPU itself. SO if there is X watts going to the CPU it doesn't mean the CPU actually needs that much power but since it is being delivered it should be called the "actual power consumption".
For example the 6000+ which needs quite too much obviously. We know now it had something to do with the different VRMs at the AM2 and AM2+ boards. The power is being delivered, meaning the power supply needs to provide this amount x. If the CPU actually needs it or not is irrelevant. Of course if I would have known from the difference between AM2 and AM2+ I would have measured it in a different way, no one is perfect and I know now how to do it better next time.
As for the Chipset it is indeed tricky since every mobo vendor has different additional chips installed that take a different amount of power. So if we do publish an article about this we will have to mention the actual manufacturer of course.
It's just funny that everyone says the results are too high but I get an Email from Taipei that my numbers are far too low...
Barack Obama - Tuesday, September 23, 2008 - link
Notch one up for the great AT team :D