The Math Behind GPU Power Consumption And PSUs

[Onus]

Love in depth articles like this. Wish there was a lot more of this. (the 2part graphics card myths one was another good one).

At the end of the day tho. HTF do we actually know what we are getting. Its been about 5 years since i bought my last PSU, but i never saw any data on what caps are in it, or what gauge wire they use etc. And i never come across power supply reviews that go into enough depth to know what you are getting.

Use the review database here:

http://www.realhardtechx.com/index_archivos/Page541.htm

The reputable review sites do a teardown of the PSU and reveal such things as capacitor brands and wire gauges.

To elaborate a little on this, IMHO the best PSU reviews were done by Gabriel Torres at HardwareSecrets. Unfortunately, he gave up his green card and returned to Brazil earlier this year, and has not done any reviews since. HardOCP and Jonnyguru also do very good ones. HardwareInsights does a nice but not quite as detailed job, but is an Australian site and what he (c.hegge) reviews may not be found in the North American market. Not quite as good but still up there are Anandtech and XbitLabs.

 

[DroopySpike]

Good info for a complex subject. The surge events from the big GPU and CPU setups can also be addressed by the end GPU/CPU designers rather than asking the ATX supply to take all the brunt. ECL/CML logic for example do not have fast transient loads. Sw techniques also can be done.

 

[f-14]

Igor Wallossek excellent article, if tuan isn't proud, i definitely am! this is the article type of toms of old earned it's reputation on. the thermal visual was quite possibly the most cloud clearing i've had just on HOW MUCH heat those areas produced...i knew they were high.... but i had figured the chipsets also produced near similar heat. it explains a whole lot of issues i've had with evga motherboards and the parts they burned out corresponding to the voltage regs having problems. the thermal imaging would probably reveal the very same thing. a very refreshing reminder as compared to power point graphs.

 

[RedJaron]

*slow clap*

Absolutely superb. Anyway these methods can be incorporated into PSU testing and reviewing in the future, or does that add too much time? Would especially love the cap spec sheet break down for future PSU reviews.

 

[jonnyguru]

A few things I want to point out.

Capacitors: While Chinese caps have gotten a bad rap in the past, they've improved vastly. And, truth be told, most "Japanese" caps are made in China. Some are made in Malaysia or the Philipines, but even companies like Nichicon, Matsushita, etc. have their largest factories in China to cater to their Chinese customers. And since 99% of PSUs are made in China... well.. do the math. ;-)

“A good example is depicted below. The manufacturer decided to go with all-black cables for aesthetics. This doesn’t matter to you or I, but the manufacturer incurs higher failure rates because the cables are harder to distinguish and identify correctly.”

Not true. The process to manufacture an all black wire cable assembly requires a little more attention to detail when it comes to matching the black wire with its associated connector terminal housing location, but cable wires are clearly marked to ensure this is done correctly. In addition, all cable sets are screened to ensure connectivity is correct. Thus, I would not agree that the failure rate is higher than a multi colored wire cable assembly process only that it might take a little longer to assemble.

“Digital PSUs are a nice addition to the market, but their switching frequencies are still quite low in most instances. This is where the use of proper capacitors on the secondary side becomes important to be able to compensate for all of the load spikes. At this point, it’s best to wait for future developments in this space, since too many unanswered questions remain and the software problems can’t be overlooked. A solid analog solution is the safe way to go for now.”

Digital PSU switching frequencies are equivalent to their analog counterparts. It looks like those voltage spikes you're getting with the graphics card loads are self induced based on the test setup, it is very likely you're picking up common mode noise. Are you using a differential mode probe?

 

[skit75]

Enjoyed this article at the end of my work day. Well done!

 

[Johan Kryger Haglert]

Jonny Guru often have written there's no problem with multi-rail designs but I haven't really knew what it was all about.

Anyway I have a EVGA SuperNova G2 750 watt which is a single rail design.

Since power demands decreased with the new graphics cards and since processors has also gone down and especially since I'll likely only use one graphics card that's massively over-powered.

But it's a single rail PSU.

Is that a problem in my case?

 

[FormatC]

@jonnyguru:
I've tried the same with another setup. The measurable "heartbeat" of the PSU is more or less very similar. About the frequencies of this digital PSUs: I've discussed it in a factory with the R&D engineers and got exactly this info. We had longer talks about all this problems around current VGA cards and they also were able to reproduce my situation in their labs.

Since VGA cards switches faster than PSUs it is in the main the job of the VGA manufacturers to implement a better filtering to smooth this idiotic spikes (but it is always a problem to place more and bigger solids below the coolers). The improvemt must be done finally on both sides, VGA and PSU, not only PSU.

The black cable thing we've discussed in a factory directly on the manual insertion line. I know that you are using black cables, but ask the girls on this line, their gaffer or the QC girls. It is not dramatic but you can register a faster fatigue and more individual errors

@Johan Kryger Haglert:
There is no problem but an really oversized PSU might be not so efficient under lower load levels or in idle.

 

[truegenius]

A table fan to cool the rig

 

[FormatC]

No. This rig is water cooler. But I need it to cool down all components faster between the tests.

 

[stunny]

Well Done, Mr. Igor Wallossek... The article was really entertaining and well written, and even f i didnt read it all (i am not that crazy, lol...), i sense your minutes, hours, days which you put in it... Congratulations, you did a big and so usefull job...

p.s. I bet sometimes you woke up in midnight, shouting "capacitors! voltage!! cables!!!"

 

[FormatC]

No. I have two small boys. The reasons why are more "natural"

 

[KaiserPhantasma]

correct me if I'm wrong but using a 760 watt PSU (AX 760 platinum... and yes this is the seasonic one) should be more than enough for a crossfire/SLI set-up as this article can even power up/use a computer and put it through testing with just a 500 watt PSU?

 

[dragosmp]

Great article, really great. In industrial equipment design we have the habit of evaluating wire gauge as function of the short-circuit for (at least) safety. Ripple current is also quite important for the losses in many equipments as obviously dirty current affects not only stability (which can be binary - on/off), but also instantaneous losses and lifetime expectancy.

An idea, for those who have ripple-related issues and don't necessarily want to invest in a new PSU, maybe you can vet it. Is it acceptable to add filtering caps on the +12V PEG, sort of PEG-to-PEG adapter with caps between the GND and +12V? Properly sized caps would cut the ripple from the GFX card, which would no longer propagate thru the PSU cables, which themselves being inductive are more sensitive the higher the frequency. I guess I'm arguing for compensating the poor GFX card board design rather than placing the onus on the PSU; when designing a PCB (as I do for other things) one can't just suppose the voltage from the power supply is perfect, constant and you can do whatever you please because the ATX standard so.

 

[dwatterworth]

Thank you very much for doing all of this work! What a huge help in understanding the fundamentals past observable build quality. Hope to read more in depth and addicting articles soon!

 

[Onus]

...
An idea, for those who have ripple-related issues and don't necessarily want to invest in a new PSU, maybe you can vet it. Is it acceptable to add filtering caps on the +12V PEG, sort of PEG-to-PEG adapter with caps between the GND and +12V? Properly sized caps would cut the ripple from the GFX card, which would no longer propagate thru the PSU cables, which themselves being inductive are more sensitive the higher the frequency. I guess I'm arguing for compensating the poor GFX card board design rather than placing the onus on the PSU; when designing a PCB (as I do for other things) one can't just suppose the voltage from the power supply is perfect, constant and you can do whatever you please because the ATX standard so.

This is what I was suggesting; that high-powered graphics cards could incorporate additional filtering. I do get to decide where my attention (the Onus) goes though :lol:, and it is indeed often on the PSU...

 

[CeruLucifus]

Terrific article. Thanks for all that hard work developing an accurate metering system.

What the PSU industry really needs is some better details to aid consumers trying to pick a model.

It would be helpful if you could give some guidance for builds that need more than 500W (multi-GPU for example). If I've read what is here correctly you presently say 1) make sure there's enough power but 2) avoid multi-rail PSUs and 3) don't get a PSU that exceeds what you need.

NoShot's example above is a good one. No doubt his system builder thought the shipped PSU was sufficient. NoShot can replace with higher capacity but how to 1) get enough, 2) make sure an individual rail won't be underpowered and 3) not get too much?

My own example: I'm doing a 2x GTX 980 SLI build right now. The cards (Asus reference model) say "Minimum 500W PSU with 38A on the 12V rail". I thought it would be easy to shop for a PSU: just to make sure combined 12V rails supply 80A and (most likely) a 1000W PSU rating. Now this article calls that selection approach into question.

 

[Ninjawithagun]

The article was well done, but in the end the conclusions drawn were too broad and speculative without any facts to back up some of them. For example, having a higher rated and higher quality power supply will negate almost if not all of the negative aspects outlined within this article. Too many people will read this article and will think that buying the absolute minimum power supply is 'good enough' for their needs and won't do any more homework than that. In reality, buying the best and highest rated power supply they can afford is the best way to go IMHO. The one thing you NEVER skimp on is the power supply. I own a Corsair AX1500i and my system only uses about 935W maximum at peak loads. So why did I buy such an 'overpowered' power supply? Simple. The power supply will degrade in overall efficiency and total output capacity over the next several years I plan to own. it. The more important reason I got it was because it is an all digital power supply that really does work as intended. It attains an average of 94% to 97% efficiency all the time and it runs absolutely silent even when under heavy load benchmarking in 3-way SLI (3 x GTX780 Gigabyte Ghz Edition cards and OCd Core i7 3930K @ 4.6Ghz 1.38V). At full load, my power supply only reaches 62% capacity load and because of that there is plenty of room for clean power on the single 12V 125A rail. Bottom line, get yourself the best power supply you can possibly afford and you won't regret it.

 

[Onus]

Ninja, I didn't quite get the same take. A lot of people are indeed inclined to buy the minimum-rated PSU, and the article explains why that may not be the best idea. Still, minimum rating wouldn't be so bad (if the quality is otherwise excellent), as long as they have enough filtering capacity to overcome those momentary spikes.
My PCs are mostly unremarkable, but I think I put my SG-650 in my will. It's "only" 80+ bronze, but I'll probably croak before it does.

 

[Eggz]

Great article!

I appreciate that someone is out there reviewing this stuff, because I could never do it. Seeing all that goes into the reviews of PSUs at Tom's is reassuring.

One a separate note, here's a few typos to fix:

(1) P.1, in text of first heading, "discovered some phenomenon" should be "phenomena"

(2) P.6, under Good Solutions and Better Solutions, "that's eliminate" should be "that'll eliminate."

(3) P.6, under Good Solutions and Better Solutions, "you or I" should be "you or me."

(4) P.9, first paragraph, capitalize after colon so "seen for what they are: a challenge . . ." should read "seen for what they are: A challenge . . . ."

 

[r0llinlacs]

When picking a power supply, or anything with a wattage rating for that matter (cough cough car amplifiers cough cough) always go by the amperage. I've browsed many PSUs that claim certain wattages, but when you go look at the amperage supplied, it doesn't add up. (IE a PSU that claims 500watts, but only has 30amps available to the 12v rails). This would be called a very overrated power supply. I'm lucky enough to have a very underrated power supply (Claims 650watts, yet has 69amps or 828 watts just on the 12v rail)

Car audio amplifiers are the biggest offenders of lying about wattage ratings and this is why the CEA wattage measurement standard was implemented (and why I always go by amperage). (IE a Boss amp claiming to do 3000watts with only 60 amps of fusing, 60A*12v = a theoretical 720watts at 100% efficiency)

Wattage ratings just cause confusion, and when people believe the ratings, it causes problems. You people told me my 650watt power supply would not be sufficient for my new R9 290X. (Probably because it says it requires a 750watt PSU) I did all the math and found I should be just fine, and I was right, I am. Seeing as how the 290X only draws an absolute maximum of 25 amps, and my PSU supplies 69A@12v, that leaves the rest of my computer with a generous 44 amps to play with, regardless of the fact my PSU claims to be only "650 watts".

When buying a PSU, your first question should always be "how many amps does it supply". When buying anything that needs powered, your first question should always be "how many amps does it draw". But don't go putting a 100amp load on a 100amp supply, as you will more than likely see severe voltage drop and horrible efficiency. Leave a little headroom on the supply side, say for example don't go over 80amp load on a 100amp supply, and you should be fine.



+3.3V@25A, +5V@25A, +12V1@22A, +12V2@22A, +12V3@25A, -12V@0.5A, +5VSB@2.5A

^ These are the most important ratings to look at

 

[Onus]

^All well and good, BUT it looks like you simply added up your +12V amperages to get a total. This almost never applies, especially on more powerful units. The +12V source has a total amperage (may be ~53A on a 650W PSU), but the individual rails have OCP set points that allow them to carry more than 1/#of_rails of that current, although the total amperage on all of them may still not exceed the rated maximum. Your amperage table should list that amount.
...which is not to say the PSU can't do more than that. For example, according to Jonnyguru's testing, my SG-650 would actually put out 894W before gracefully shutting down.

 

[Johan Kryger Haglert]

@Johan Kryger Haglert:
There is no problem but an really oversized PSU might be not so efficient under lower load levels or in idle.

True.

Guess I'd figured the power demands had always gone up. I mean my previous PSU was 360 watt and I hadn't had a problem with that, and graphics cards have two slot large coolers and lots more material and bigger fans now.

AND THEN THEY CHANGED!

I really didn't needed two cards from the beginning.

http://www.jonnyguru.com/modules.php?name=NDReviews&op=Story3&reid=380
82.5-88.5% at ~150 watt there depending on the test.

But I guess a i5 4430 / 4690K / 5820K and a GTX 970 will not use that on idle. Under 100?

 

[Johan Kryger Haglert]

Night owl: Didn't he answer that? To begin with there's a difference in two GTX 750 Ti and two 290x2. And wasn't there some part about the motherboard turning it all off due to poor power or whatever? Maybe not whatever it would work but yeah. Maybe they have a generic answer, maybe being more specific about both cards and equality/price level of PSU would help?

 

[beetlejuicegr]

I know this could be a noob question but i will do it. So, is it better to use 2 PSUs of 500watts, giving one 8pin PCI-E cable each to a hungry GPU each and splitting the rest of the PC power needs than having a 1000watt of same quality PSU? Is that it?