And a second attempt at this conversation ends fruitlessly.

Michapolys

rjohnson11
We've had some complaints about this thread. I ask that everyone please respect each other's opinions even if you don't agree with them. 

Fine by me. Care to share what was the complaint?

He's a moderator. He doesn't have to share squat with someone like you.

Michapolys

kevinc313
Here's a quick discussion about switching in computer PSU's and the associated datasheets, I'm not seeing anything about switching in the multiple 100's of hz range.  They are variable in different ways and for low power but sounds like they always run at many khz.
 
https://www.tomshardware....plies-101,4193-13.html

The information there briefly mentions the part I am referring to.

A more detailed explanation of what I am talking about can be found here:

https://www.powersystemsd...e-edged-sword/31/14595

Figure 2 there specifically shows the behavior I am talking about, where all cycles are spent in bursts and then you are left with exponentially increased dead time where the effective frequency zeroes out. Your capacitance bank is meant to handle potential increased needs in current during that time.

That's for a low-power external power supply. I have my doubts that a PC power supply would have such a light load to pause for such long durations.

What I think needs attention is the condition of failure. The condition of failure is that the GPU's fuse blows while the rest of the system continues to operate properly. I haven't seen your theory address or explain that.

Michapolys

kevinc313
Here's a quick discussion about switching in computer PSU's and the associated datasheets, I'm not seeing anything about switching in the multiple 100's of hz range.  They are variable in different ways and for low power but sounds like they always run at many khz.
 
https://www.tomshardware....plies-101,4193-13.html

The information there briefly mentions the part I am referring to.

A more detailed explanation of what I am talking about can be found here:

https://www.powersystemsd...e-edged-sword/31/14595

Figure 2 there specifically shows the behavior I am talking about, where all cycles are spent in bursts and then you are left with exponentially increased dead time where the effective frequency zeroes out. Your capacitance bank is meant to handle potential increased needs in current during that time.

Very interesting.  Are you saying you've personally observed this behavior with a modern ATX power supply using an oscilloscope?  Sounds like an improper implementation of the power saving features these controllers come with.

Michapolys

kevinc313
Very interesting.  Are you saying you've personally observed this behavior with a modern ATX power supply using an oscilloscope?  Sounds like an improper implementation of the power saving features these controllers come with.

I was interested a few years ago to see why some PSUs would trip their protections with AMD Vega GPUs and had a look around for transient specs.
Came to the conclusion that the ATX specification being easy to misinterpret was to blame.
Long story short, most ATX PSUs and older GPUs lean closer to the minimum required power requirements, while more modern GPUs lean more towards the recommended requirements.

It is a pity too because modern ATX PSUs are meant to be able to reach high switching frequencies so they should be fine as long as you put an adequate minimum load to them.

My guess is that the need to support the recent energy saving regulations and higher grade 80plus badges (marketing) is what resulted in this. That was the conclusion I ended up with.

Seems like an interesting theory.  Did you actually put a scope on a PSU?  Is there a way to determine, besides direct testing, what minimum load would keep a PSU out of this low power state, for say a certain rating and platform?

Michapolys

ty_ger07
That's for a low-power external power supply. I have my doubts that a PC power supply would have such a light load to pause for such long durations.

What I think needs attention is the condition of failure. The condition of failure is that the GPU's fuse blows while the rest of the system continues to operate properly. I haven't seen your theory address or explain that.

It is a typical example of how a modern LLC resonant based topology works. The specific frequencies used are up to each specific design and what are your target specs. Generally speaking, frequency extremes are avoided from a point onwards on ATX LLC implementations due to cost and/or compatibility concerns. The whole idea of my post was to get the point across.

The fuse you are referring to seemed to be the 20A kind, meaning about 40A needed to blow it (doubt it is higher due to safety concerns, but it could be lower if they are conservative).
Your load does not care how it gets its power, as long as it gets it as fast as possible (path of least resistance). The protections do, if they exist somewhere (either on the load, psu, mains, etc.) and your fuses count as protections. Your load itself however, does not care what you feed it.

I did explain the voltage/current/capacitance relation on my original post about the potential problem workarounds, so you can more or less tell what your voltage was at that point. Blown fuse means that the rest protections did not trigger in time.

So, you are theorizing that amperage increases because the voltage drops. How far does the 12v supply need to drop while the video card VRM buck converter continues to attempt to operate at the same power draw? The 12v supply needs to drop to around 3 volts? Do you think the motherboard, video card, and PSU undervoltage protection won't kick in way before then? Do you think the rest of the computer will continue to run, with the video card fans running at 100%, CPU and memory receiving plenty of voltage from the same exact power source to continue operating normally, and the operating system continuing to operate with game audio continuing to play? How does the theory fit the observation?

I feel like I am asking the same questions I asked 5 weeks ago.

The problem for you is that even if you hook the GPU directly up to a DC battery, you can still demonstrate that it will draw power spikes. The power spikes are not a function of the power supply. If you define your power limit as an average, transient spikes are assured. If you determine that the video card's overcurrent protection is set higher than the VRM's designed limit, what is the result you should expect?

Why are we out in left field hypothesizing about something which doesn't match the observation?

kevinc313Fair enough.  Most PSU's don't hit their nominal efficiency until about 20% load, where I assume they are operating normally.  You can see a blip in this 1000w G6 plot around 65w which may be a switching mode change out of low power mode.  That's above idle power draw.
 

I would not look at the G6 for reference, as Seasonic made that platform specifically to save face due to their previous modern platforms having issues with transients. I get your point though.

A more obvious case of the blip, with more severe increase in efficiency at lower loads is on one of their more efficient implementations, like here for example:

Can we get back on track?

So, you are theorizing that amperage increases because the voltage drops. How far does the 12v supply need to drop while the video card VRM buck converter continues to attempt to operate at the same power draw? The 12v supply needs to drop to around 3 volts? Do you think the motherboard, video card, and PSU undervoltage protection won't kick in way before then? Do you think the rest of the computer will continue to run, with the video card fans running at 100%, CPU and memory receiving plenty of voltage from the same exact power source to continue operating normally, and the operating system continuing to operate with game audio continuing to play? How does the theory fit the observation?

I feel like I am asking the same questions I asked 5 weeks ago.

The problem for you is that even if you hook the GPU directly up to a DC battery, you can still demonstrate that it will draw power spikes. The power spikes are not a function of the power supply. If you define your power limit as an average, transient spikes are assured. If you determine that the video card's overcurrent protection is set higher than the VRM's designed limit, what is the result you should expect? Why was the overcurrent protection set so high? To avoid black screen events which plagued previous generations, NVIDIA set the overcurrent protection higher than practical. Rather than protect the video card, NVIDIA decided that it was more important that users not be annoyed by intermittent black screen crashes. NVIDIA hopes that the video card VRM will survive the transients, but then when EVGA releases the FTW3 with a "XOC" BIOS that everyone uses (and then starts selling the card with the XOC BIOS included), EVGA pushes the NVIDIA reference VRM too far so that transient spikes start killing that particular model. When you raise your average power limit, your transient spikes get larger.

Why are we out in left field hypothesizing about something which doesn't match the observation?

According to your theory, why does this problem not affect Founders Edition or Kingpin cards?

kevinc313
Because it's nice to contemplate a new theory vs. repeating the same common knowledge.
 
BTW one of the power shutoffs I had on my 1300w Prime Gold with a 3080 Ti FTW3 was going from a cutscene to gameplay max settings 4K RTX in Control, I would suspect there was fairly low power in the transition.

I keep hearing about "theory" but really, it is just information, most of it being available to have a look at yourself, if you look at the related documents.

One is able to pinpoint exactly what is going on by looking at the voltage oscillation of the output under the problematic conditions.

kevinc313
Yes, it's a theory until a qualified person puts a scope on a PSU.  

They did. I already said that earlier. You can ask PSU OEMs for data yourself too. The specs for various resonant controllers used on ATX PSUs are also available to look at, so are various versions of the ATX spec.

kevinc313
Thanks, please post them or link to them.

Now you are asking too much. If something was too hard to get a hold of, I could lend a hand. It is not.

You are perfectly capable of taking a hold of all the related documentation yourself.

kevinc313
YOU owe ME.

No.

kevinc313

ty_ger07Why are we out in left field hypothesizing about something which doesn't match the observation?

Because it's nice to contemplate a new theory vs. repeating the same common knowledge.
 
BTW one of the power shutoffs I had on my 1300w Prime Gold with a 3080 Ti FTW3 was going from a cutscene to gameplay max settings 4K RTX in Control, I would suspect there was fairly low power in the transition.

But it's still not what is killing FTW3s. Did your FTW3 die? That is the subject of this thread. It can't fit the observations of FTW3 deaths.

Michapolys

kevinc313
YOU owe ME.

No.

i'm surprised you've entertained them this long honestly. you should get a big blue ribbon imo. internets is dark and full of low info trolls

Michapolys
I keep hearing about "theory" but really, it is just information, most of it being available to have a look at yourself, if you look at the related documents.

The "theory" isn't about how PSUs work. The "theory" is that what you are talking about has any bearing on the reason why FTW3s fail. So far, you haven't provided any evidence of it and have skirted around every one of my questions regarding applicability to the subject at hand.