Concerning GPU Transient power consumption problem

[deadpoor]

This week FCP published a shocking article detailing how much peak power consumption of gpu's is and the problems that can arise that most people currently don't realize. The article is very detailed and compelling,It’s mind blow for me and Chines PC enthusiast, so I've included a short excerpt for discussion.

 

3-2. Graphics Card Load Characteristics

Since games nowadays run at dozens or even hundreds of frames, to analyze the load characteristics of the graphics card you need to measure the load of the graphics card in milliseconds or even microseconds for each frame.

Example: 2070 super running BF5, locked at 60 frames, each frame rendering cycle 16.7ms. in a rendering cycle, the graphics card has two states, rendering and resting, the peak power consumption is about 430W, 2-3 times the EDP, the power consumption is reduced to 80w when resting, and the maximum average power consumption is 200w. 

3-4. CPU and graphics card peak power phase

LinX+Furmark benchmark with 1036W instant peak power consumption.

 

3-5, “Rumored” official Nvidia graphics card peak power specifications

Maximum total power consumption is 77A/924W/1ms

 

3-6, a bunch of graphics cards peak power consumption measurement

(test platform at origin article)

 

 

 

3-7. Analysis of peak power consumption of the graphics card

When the peak power/TDP ratio(last column in list) is too large, the Piezoelectricity effect occurs due to frequent switching of transistors to control the total power consumption.

 

5-4, OCP/OPP overload protection mechanism

OCP (Over Current Protection) is a design mandated by Intel, the purpose of which is to cut off the output of the power supply when the output current is too high to maintain system safety, and interfaces, cables and other components should not be melted or damaged. According to the design of different manufacturers, the current power supply's overall OPP may be between 115% and 160% of the rated power, different power supplies can withstand overload time is also different, there will not be a long overload, usually a few hundred microseconds or a few milliseconds.

Power overload shutdown case, peak power consumption 1124W, pulse width 500uS

 

Overload test for different models

 

model	FCPG rate	OPP test	overload behave
恩杰NZXT E850	A+	~140%	normal gaming
华硕ROG Strix 750W Gold	A	1044.53W (139.27%)	shutdown protection
超频三PCCooler 七防芯GI-P850	A	1152.12W (135.5%)@560μs	shutdown protection
威刚 ADATA XPG CORE REACTOR 650W	A+	905.5W (139.3%)@47ms	normal gaming,12V undershoot risk 9.87V
爱国者 Aigo ES650	B	923.3W (142%)@7ms	normal gaming,12V overshoot risk12.9V
酷冷至尊 CoolerMaster V650	C	793.3W (122%)@4ms	12V overshoot，shutdown protection
全汉 FSP Hydro GE 650	FAIL	747.4W (115%)@18ms	no protection,12V out of control 10.68~13.44V
恩杰 NZXT C650	B	916.2W (141%)@3.02	shutdown protection
海韵 Seasonic FOCUS GX-650	B	905.6W(139.3%)@3ms	shutdown protection
振华 SuperFlower Leadex Gold 650	A	994.2W (153%)@12ms	PSU Abnormal noise
富钧 Xigmatek Harrier HA650	D	988.4W (152.1%)@8ms	shutdown protection
海盗船RM650x 2018	A	875W (134.7%)@3.14ms	shutdown protection
海韵Seasonic FOCUS SGX-650	A	909.4W (139.9%)@4.78ms	shutdown protection
海盗船Corsair SF750	A+	1011.09W (134.8%)@1.26ms	normal gaming,12V overshoot risk12.6V
海韵 Seasonic FOCUS GX-850	A+	1285.66W (151.25%)@~2.5ms	normal gaming
振华 SuperFlower Leadex III Gold 650 ARGB	A+	924.65W (142.2%)@17ms	normal gaming
酷冷至尊 CoolerMaster MWE Gold V2 650W	A	913.15W (140.48%)@2.54ms	shutdown protection
海韵 Seasonic CONNECT 1.0	A	1,009.33W (134.6%)@~2.5ms	shutdown protection
酷冷至尊 CoolerMaster V850 SFX	A+	1225.22W (144.14%)@8.5ms	normal gaming,12V overshoot risk12.7V
恩杰NZXT C850	A+	1275.62W (150.07%)@2ms	normal gaming
九州风神DeepCool DQ850M-V2L	A	1096.79W (129.79%)@2.63ms	normal gaming, OPP low

 

To sum up, if the power supply overload protection time is longer than the peak power consumption time, you can survive the peak power consumption, but the power supply itself will increase the heat of the components, affecting the life of components, followed by an increase in the risk of “blowing up the machine”, the protection trigger time is too long, encounter really want to protect the moment, the protection function will appear sluggish, often resulting in more serious consequences, the power supply output voltage regulation is also easy to run out of control.

 

8-3, the beginning of a power plant lying shot

At the end of 2018, Techpowerup pointed out that the seasonic power supply has a serious failure, FOCUS+ 550 and ASUS graphics card or vega56/64 with will appear black screen, shutdown. seasonic official reply, after testing vega 56 crossfire got 102A/10ms peak power consumption, due to the power supply design tends to protect, relatively easy to shutdown, after 2018, the power supply has modified the overload protection sensitivity.

This article was originally published by @fcpowerup at FCPOWERUP.COM. Due to the length, I only relayed a very small part of which I think is important for now, I have two questions 1. Is the peak power consumption of the graphics card really far more than TDP as stated in this article. 2. Is the peak power consumption really has such a great impact on the stability of the power supply and system. I hope LTT can verify it. If it's true as stated in this article, then the recent power supply buying guides released by gamersnexus are severely underestimating, so I hope won't be misleading to gamers. And the PCAT  looks like another green salt sprinkled by NVIDIA.

Sorry for poorly translated article, I’m not a native English speaker, but I will be more than happy to answer any questions you may have.

[PeachGr]

I m not an expert, also I can't understand everything. I m having a r7_3800x and a 2060 super and while folding (maximum load) I am at 330w and that includes a few LED from 4 fans+ keyboard and mouse LED+ everything else running on the PC.

I m guessing that the "over the limit" you mentioned above, occurs to some overclock? Because I cannot believe that my GPU can drain that much power alone. 

But as I said I don't understand everything

Also ltt just released a related video

[Bombastinator]

So what you’re saying is there is a claim that the wattage ratings of some GPUs are false, and their peak power draw is vastly higher than their rated wattages and a 2060 can draw over 1000w for..microseconds?  My memory is this sort of thing is generally taken care of with capacitors.  I notice they were testing 5700xts with a 500w PSU where the blanket recommendation for a 5700xt is 650w

 

what is FCP?

 

This was originally a tweet?  There’s a lot of fairly professional looking photography for a tweet though I suppose phones have gotten a lot better at such things.  The cluttered area in 3.6 would have been very hard to light.  Also the very deep field of view suggests a low f-stop. 20 years ago such a set of pictures would have cost many thousands of dollars to produce.  My experience with such things is very out of date though.  The references to diagram numbers also seems strange. Not exactly a tweet sort of thing. Also the sheer cost of the diagnostic equipment on the rack shelving would be pretty high.

 

I do not trust twitter for facts much.  Anyone can put anything there. Doesn’t make it automatically false though.  I lack knowledge.  I’m curious what more knowledgeable people have to say about this. 

 

 

[deadpoor]

9 hours ago, PeachGr said:

I m not an expert, also I can't understand everything. I m having a r7_3800x and a 2060 super and while folding (maximum load) I am at 330w and that includes a few LED from 4 fans+ keyboard and mouse LED+ everything else running on the PC.

I m guessing that the "over the limit" you mentioned above, occurs to some overclock? Because I cannot believe that my GPU can drain that much power alone. 

But as I said I don't understand everything

Also ltt just released a related video

yes I saw the PCAT video, that's the reason I regist and post here, https://www.igorslab.de/en/nvidia-pcat-in-test-convenient-power-acquisition-measurement-and-data-acquisition-from-frame-time-to-system-data/  igor'slab posted a review of it yesterday, that's a big different between PCAT and oscilloscope.

this article is not about overclock, in first picture,when render one frame in 16.7ms, GPU have two stats, work and rest. average power consumption is just 200w. but max and min is 430w and 80w, it's kind big for PSU to supply. if PSU have longer overload protection time then it's function normal.

[deadpoor]

8 hours ago, Bombastinator said:

So what you’re saying is there is a claim that the wattage ratings of some GPUs are false, and their peak power draw is vastly higher than their rated wattages and a 2060 can draw over 1000w for..microseconds?  My memory is this sort of thing is generally taken care of with capacitors.  I notice they were testing 5700xts with a 500w PSU where the blanket recommendation for a 5700xt is 650w

 

what is FCP?

 

This was originally a tweet?  There’s a lot of fairly professional looking photography for a tweet though I suppose phones have gotten a lot better at such things.  The cluttered area in 3.6 would have been very hard to light.  Also the very deep field of view suggests a low f-stop. 20 years ago such a set of pictures would have cost many thousands of dollars to produce.  My experience with such things is very out of date though.  The references to diagram numbers also seems strange. Not exactly a tweet sort of thing. Also the sheer cost of the diagnostic equipment on the rack shelving would be pretty high.

 

I do not trust twitter for facts much.  Anyone can put anything there. Doesn’t make it automatically false though.  I lack knowledge.  I’m curious what more knowledgeable people have to say about this. 

 

 

origin article at https://www.fcpowerup.com/end-gpu-spike-peakpower/

I'm confused, where did you see 2060 with 1000w, in 3.6 it says 421w. In origin post 7-6 it stats capacitors will be too much or too expensive to handle this, and will causing other problems.

 

FCPowrup is a professional PSU review site in china, it post on Weibo and got thounds retweet, I post author‘s tweet for convenience. origin post is too long just like a essay, he stats he start this project at 2018 after seasonic focus+550 failure, you could see the 2080 and vega in 3-6 for relative timing.

 

https://www.fcpowerup.com/how-do-i-test-power-supplies/ In another article author describes his PSU review methods in detail, l'm not professional but hell they look like expensive. he even criticize igorslab's method not accurate enough.

 

2*Chroma 8000 Test System

Tektronix MDO3014  Oscilloscope

Tektronix TDS3014 Oscilloscope

SW826 Non-contact Laser Velocimeter

FILR One Pro

 

I’m also curious too, I find that hard to belive Gamersnexus doesn't realize about this and think it's waste money on PSU, I'd like to see third party like LTT jump in and verify those result.

 

 

[Mark Kaine]

The thing that doesn't make sense to me, most people don't use 1000+ w psus , they often use a little lower than recommended' even. 

 

So how can a 500w psu draw  for example ~700w?

 

Maybe it can but it doesn't make much sense to me... If you know what I mean...

 

 

Edit: ok I read up, so it's working as intended (since many years) that's why psus have overload protection.

 

What is the problem?

 

 

[Bombastinator]

3 hours ago, deadpoor said:

origin article at https://www.fcpowerup.com/end-gpu-spike-peakpower/

I'm confused, where did you see 2060 with 1000w, in 3.6 it says 421w. In origin post 7-6 it stats capacitors will be too much or too expensive to handle this, and will causing other problems.

 

FCPowrup is a professional PSU review site in china, it post on Weibo and got thounds retweet, I post author‘s tweet for convenience. origin post is too long just like a essay, he stats he start this project at 2018 after seasonic focus+550 failure, you could see the 2080 and vega in 3-6 for relative timing.

 

https://www.fcpowerup.com/how-do-i-test-power-supplies/ In another article author describes his PSU review methods in detail, l'm not professional but hell they look like expensive. he even criticize igorslab's method not accurate enough.

 

2*Chroma 8000 Test System

Tektronix MDO3014  Oscilloscope

Tektronix TDS3014 Oscilloscope

SW826 Non-contact Laser Velocimeter

FILR One Pro

 

I’m also curious too, I find that hard to belive Gamersnexus doesn't realize about this and think it's waste money on PSU, I'd like to see third party like LTT jump in and verify those result.

 

 

The number was 1036w I believe. It was iirc in the first paragraph.  It was sort of the whole basis for the concept it seemed to me at least, part of the problem.  I’m not competent to critique the electronics stuff and even my photography experience is so out of date it may no longer be useful. I do want to see what someone that is competent thinks.  The problem with any research like this is it can miss things. That’s the way research works, and therefore why peer review exists.  I am not a peer.  I cannot review.  

[deadpoor]

4 hours ago, Mark Kaine said:

The thing that doesn't make sense to me, most people don't use 1000+ w psus , they often use a little lower than recommended' even. 

 

So how can a 500w psu draw  for example ~700w?

 

Maybe it can but it doesn't make much sense to me... If you know what I mean...

 

 

Edit: ok I read up, so it's working as intended (since many years) that's why psus have overload protection.

 

What is the problem?

 

 

problem is we don't know how long and how much will GPU peak, we also don't know when a PSU will trigger protection, GPU manufacturers didn't provide those data, PSU manufacturers didn't have guideline of how to engage OPP.  Reviewer didn't have accurate method to measure it. 

[Mark Kaine]

3 hours ago, deadpoor said:

problem is we don't know how long and how much will GPU peak, we also don't know when a PSU will trigger protection, GPU manufacturers didn't provide those data, PSU manufacturers didn't have guideline of how to engage OPP.  Reviewer didn't have accurate method to measure it. 

Yes ok I see it is an issue - though not a new one, it would probably help if PSU manufacturer had better data for sure. It's not a "big" issue either, though I suppose that's one of the reasons cheap PSUs are bad as they likely won't be tested properly.

 

So yes, this is an interesting find for sure that should be talked more about.

 

 

 

 

[deadpoor]

On 9/9/2020 at 7:35 AM, Bombastinator said:

So what you’re saying is there is a claim that the wattage ratings of some GPUs are false, and their peak power draw is vastly higher than their rated wattages and a 2060 can draw over 1000w for..microseconds?  My memory is this sort of thing is generally taken care of with capacitors.  I notice they were testing 5700xts with a 500w PSU where the blanket recommendation for a 5700xt is 650w

 

what is FCP?

 

This was originally a tweet?  There’s a lot of fairly professional looking photography for a tweet though I suppose phones have gotten a lot better at such things.  The cluttered area in 3.6 would have been very hard to light.  Also the very deep field of view suggests a low f-stop. 20 years ago such a set of pictures would have cost many thousands of dollars to produce.  My experience with such things is very out of date though.  The references to diagram numbers also seems strange. Not exactly a tweet sort of thing. Also the sheer cost of the diagnostic equipment on the rack shelving would be pretty high.

 

I do not trust twitter for facts much.  Anyone can put anything there. Doesn’t make it automatically false though.  I lack knowledge.  I’m curious what more knowledgeable people have to say about this. 

 

 

well, the more knowledgeable future steve said the same thing two year after, better late than never...I guess.