One thing that should be added to the tests - noisy power consumption. Modern graphics cards have current loadings that vary wildly from millisecond to millisecond but almost all tests are carried out with smooth loads. For a test that more closely reproduces the load in a PC, the load current of a modern graphics card playing a demanding game should be recorded and that same load profile (scaled according to the PSU rating) should be applied to the PSU. The voltage regulation for a static load is likely to be far better than for a load that varies from 10% to 90% then back again many times a second. Varying loads can also put a higher strain on the PSU - a PSU that can provide 700W as a steady load may fail at a highly dynamic load that never exceeds 650W.
I agree with this, but unfortunately executing that sort of test is going to be pretty difficult. First of all you are going to need to put a shunt and an oscilloscope on the rails of the video card (much like tom's hardware did on one of their GPU power articles -- they actually intercepted the power coming in through the pcie slot + the power from the extra connectors). Then you are pretty much going to need to build something to 'replay' that power trace. Probably something that has a MCU and a TON of FET's on a heatsink, and the MCU could switch them on and off. As far as I know there aren't really any devices out there that can replay back a power useage recording like that...
Hrmm, interesting idea though. Would be a fun little project to build that dynamic load though.
If you asked Tom's Hardware they might give you the data from their recording - save you from having to do it yourself. The actual dynamic load is basically a (very) heavy duty D/A, seven bit accuracy and a 10kHz sample rate would suffice. (A Raspberry Pi could drive it or if you have an old PC around then the D/A could even be driven from the printer port!!!)
For component validation, it's much more common to apply a regular waveform, such as a 100 Hz trapezoidal waveform from 10 to 90%, than to try to replay a recording. That is done in some tests, but (speaking as a controls engineer in the test and measurement industry) the hard part of testing is not applying the test load, it's quantifying the results. Some (relatively) simple math will tell you the amplitude of the error if you have a regular waveform and an oscilloscope trace. It's much more difficult to look at a half hour of data responding to what might as well be random stimuli at 200 MHz and make any sense of what happened. The only time that you'd really want to do that is if you're measuring something a layer removed from the stimulus - heat generated, total power consumed, and so on. "Average" or "Worst" error level wouldn't be very useful.
It would be really cool if you could complement this story with a few measurements of power usage per rail of a few of the anandtech testbeds. Just to put this in perspective.
Heck, even dual-GPU enthusiast machines don't need more than 700W-800W, tops. 1kW+ PSUs are only of interest to those running triple or quad GPUs which is a very tiny minority of a minority.
Generally they don't, and ultimately we're headed down a more efficient road... Dual R9 290s would put quite a max load on a 700W unit tho... I ended up upgrading the 750W unit I happily used with my 2x 6950 because of that, if only I'd seen the GTX 970 coming. :p
I wouldn't be too quick to assume that because nVidia's used it's advantage in the 700/900 series GPUs to run at cooler default settings (quieter and potentially lower failure rates). AMD is currently cutting prices and throwing more power at their chips for a given level of performance to stay competitive. A few generations back when the 4xx series were dogs nVidia did the same; the next time they have the weaker chip I'd expect them to do the same.
My first thought reading this article was: mighty bullshit, the guidelines don't make sense for nonsense PSU so let's change the guidelines instead of focusing on no-nonsense PSUs.
The issue is more dramatic in higher capacity designs; but a lot of newer mid-range PSUs have the same issue. I know I've seen PSUs at least as small as 550W that used the same basic design as the monsters: They can generate their entire rated output on 12V and have only 100W available for 3.3/5V rails. If they kept the same 100W for legacy rails everything for 12V; even a 300W unit would be slightly over using the derating formula given in the article. If, as seems reasonable, they used a smaller legacy supply circuit it would likely get much worse than the situation with the 1200W unit: The derating formula assumes that all rails are loaded equally, as a result reducing the max number of amps on the 3.3/5v rails from 20 to 10 increases the calculated per rail load from 64 to 78%; which works to 130W combined instead of 106. Assuming the max power was cut to 50W to go along with the 50% cut in amps, you're be running the legacy rails at 2.6x rated capacity. That's a formula for a smoke test, not just an emergency shutdown.
Agreed! We need more reviews of normal PSUs. Sadly, updating the test methodology only makes sense if we're going to see EVEN MORE reviews of huge PSUs.
I would recommend that instead of 500 watts you look at what Intel or the MB manufacturer for a dual socket server recommends and use that instead. And just use the original derating formula but calculate it as if it was that lower wattage.
I would prefer to see more tests of realistically-sized PSUs, not the kilowatt-plus behemoths that are impressive, but not really relevant to most users.
Sure, I could find tests of them elsewhere, but AT has the only tests that are really worth a shit. I've actually seen tests of PSUs that included game framerates in their "results". It's just disappointing that I can't find AT benches of any PSU I'm considering for a build.
You should look at more regular and older sites like jonnyguru.com (the original pioneers for many thanks to their antics of exploding shit-tier crap), techpowerup.com, and thanks to crmaris from TPU writing for them, Tom's hardware of all places...
Just here to add to the chorus of voices calling for reviews of PSUs in the wattage/price range that most people actually buy. 1kw+ units are cool to look at, but I've never used one in a build. It definitely seems like too much time is spent on high end stuff (not only PSUs) that only a small number will buy.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
18 Comments
Back to Article
Duncan Macdonald - Monday, May 4, 2015 - link
One thing that should be added to the tests - noisy power consumption. Modern graphics cards have current loadings that vary wildly from millisecond to millisecond but almost all tests are carried out with smooth loads. For a test that more closely reproduces the load in a PC, the load current of a modern graphics card playing a demanding game should be recorded and that same load profile (scaled according to the PSU rating) should be applied to the PSU. The voltage regulation for a static load is likely to be far better than for a load that varies from 10% to 90% then back again many times a second. Varying loads can also put a higher strain on the PSU - a PSU that can provide 700W as a steady load may fail at a highly dynamic load that never exceeds 650W.extide - Monday, May 4, 2015 - link
I agree with this, but unfortunately executing that sort of test is going to be pretty difficult. First of all you are going to need to put a shunt and an oscilloscope on the rails of the video card (much like tom's hardware did on one of their GPU power articles -- they actually intercepted the power coming in through the pcie slot + the power from the extra connectors). Then you are pretty much going to need to build something to 'replay' that power trace. Probably something that has a MCU and a TON of FET's on a heatsink, and the MCU could switch them on and off. As far as I know there aren't really any devices out there that can replay back a power useage recording like that...Hrmm, interesting idea though. Would be a fun little project to build that dynamic load though.
Duncan Macdonald - Monday, May 4, 2015 - link
If you asked Tom's Hardware they might give you the data from their recording - save you from having to do it yourself. The actual dynamic load is basically a (very) heavy duty D/A, seven bit accuracy and a 10kHz sample rate would suffice.(A Raspberry Pi could drive it or if you have an old PC around then the D/A could even be driven from the printer port!!!)
TurboTastic - Tuesday, May 5, 2015 - link
As described in the "How We Test PSUs 2014" article at http://anandtech.com/show/7820/how-we-test-psus-20... , they have several Maynuo M9714 programmable loads. These can do sinusoidal, trapezoidal, or other regular waveforms, change between two states on an external trigger, or be run (slowly) over GPIB with a preprogrammed array of data (product manual: http://www.maynuo.com/downloadfile/201001195766356...For component validation, it's much more common to apply a regular waveform, such as a 100 Hz trapezoidal waveform from 10 to 90%, than to try to replay a recording. That is done in some tests, but (speaking as a controls engineer in the test and measurement industry) the hard part of testing is not applying the test load, it's quantifying the results. Some (relatively) simple math will tell you the amplitude of the error if you have a regular waveform and an oscilloscope trace. It's much more difficult to look at a half hour of data responding to what might as well be random stimuli at 200 MHz and make any sense of what happened. The only time that you'd really want to do that is if you're measuring something a layer removed from the stimulus - heat generated, total power consumed, and so on. "Average" or "Worst" error level wouldn't be very useful.
jann5s - Monday, May 4, 2015 - link
It would be really cool if you could complement this story with a few measurements of power usage per rail of a few of the anandtech testbeds. Just to put this in perspective.ImSpartacus - Monday, May 4, 2015 - link
Can we just have reviews of more modest psus?As a consumer, it doesn't help me to read a review of a gigantic psu. I don't really about anything more than ~600W.
jann5s - Monday, May 4, 2015 - link
+1r3loaded - Monday, May 4, 2015 - link
Heck, even dual-GPU enthusiast machines don't need more than 700W-800W, tops. 1kW+ PSUs are only of interest to those running triple or quad GPUs which is a very tiny minority of a minority.Impulses - Monday, May 4, 2015 - link
Generally they don't, and ultimately we're headed down a more efficient road... Dual R9 290s would put quite a max load on a 700W unit tho... I ended up upgrading the 750W unit I happily used with my 2x 6950 because of that, if only I'd seen the GTX 970 coming. :pDanNeely - Monday, May 4, 2015 - link
I wouldn't be too quick to assume that because nVidia's used it's advantage in the 700/900 series GPUs to run at cooler default settings (quieter and potentially lower failure rates). AMD is currently cutting prices and throwing more power at their chips for a given level of performance to stay competitive. A few generations back when the 4xx series were dogs nVidia did the same; the next time they have the weaker chip I'd expect them to do the same.Daniel Egger - Monday, May 4, 2015 - link
My first thought reading this article was: mighty bullshit, the guidelines don't make sense for nonsense PSU so let's change the guidelines instead of focusing on no-nonsense PSUs.Definitely +1 from me as well.
DanNeely - Monday, May 4, 2015 - link
The issue is more dramatic in higher capacity designs; but a lot of newer mid-range PSUs have the same issue. I know I've seen PSUs at least as small as 550W that used the same basic design as the monsters: They can generate their entire rated output on 12V and have only 100W available for 3.3/5V rails. If they kept the same 100W for legacy rails everything for 12V; even a 300W unit would be slightly over using the derating formula given in the article. If, as seems reasonable, they used a smaller legacy supply circuit it would likely get much worse than the situation with the 1200W unit: The derating formula assumes that all rails are loaded equally, as a result reducing the max number of amps on the 3.3/5v rails from 20 to 10 increases the calculated per rail load from 64 to 78%; which works to 130W combined instead of 106. Assuming the max power was cut to 50W to go along with the 50% cut in amps, you're be running the legacy rails at 2.6x rated capacity. That's a formula for a smoke test, not just an emergency shutdown.TurboTastic - Tuesday, May 5, 2015 - link
Agreed! We need more reviews of normal PSUs. Sadly, updating the test methodology only makes sense if we're going to see EVEN MORE reviews of huge PSUs.toyotabedzrock - Monday, May 4, 2015 - link
I would recommend that instead of 500 watts you look at what Intel or the MB manufacturer for a dual socket server recommends and use that instead. And just use the original derating formula but calculate it as if it was that lower wattage.makerofthegames - Monday, May 4, 2015 - link
I would prefer to see more tests of realistically-sized PSUs, not the kilowatt-plus behemoths that are impressive, but not really relevant to most users.Sure, I could find tests of them elsewhere, but AT has the only tests that are really worth a shit. I've actually seen tests of PSUs that included game framerates in their "results". It's just disappointing that I can't find AT benches of any PSU I'm considering for a build.
ZeDestructor - Monday, May 4, 2015 - link
You should look at more regular and older sites like jonnyguru.com (the original pioneers for many thanks to their antics of exploding shit-tier crap), techpowerup.com, and thanks to crmaris from TPU writing for them, Tom's hardware of all places...Dorek - Tuesday, May 5, 2015 - link
You should read the PSU reviews at jonnyguru.com.Salvor - Tuesday, May 5, 2015 - link
Just here to add to the chorus of voices calling for reviews of PSUs in the wattage/price range that most people actually buy. 1kw+ units are cool to look at, but I've never used one in a build. It definitely seems like too much time is spent on high end stuff (not only PSUs) that only a small number will buy.