RTX 5090 FE Molten 12VHPWR

serpretetsky said:
I think its fine as a single thread. Its not like we're going to do failure rate analysis anyways. So why divide it. Besides erek posting i dont know what is left to discuss in this thread that would warrant dividing it. Its mostly focused on the 12vhpwr connector regardless which gpu.
Click to expand...
Funny thing is the 8 pin melting at the PSU side using a 12vhpwr adapter is 100% due to the lack of current balancing on the 12vhpwr side.
 
Tsumi said:
Funny thing is the 8 pin melting at the PSU side using a 12vhpwr adapter is 100% due to the lack of current balancing on the 12vhpwr side.
Click to expand...
When I tried to Google or even ask AI to find examples of melted 8 pin connectors, it was like 90% of the time a 12vhpwr adapter when you clicked and took a closer look.

It's very damning, lol.
 
kalston said:
When I tried to Google or even ask AI to find examples of melted 8 pin connectors, it was like 90% of the time a 12vhpwr adapter when you clicked and took a closer look.

It's very damning, lol.
Click to expand...
You google search must be very different from mine:
"8 pin melted GPU power"

Care to document your claims?
1794060026635.png
 
Last edited:
Tsumi said:
Funny thing is the 8 pin melting at the PSU side using a 12vhpwr adapter is 100% due to the lack of current balancing on the 12vhpwr side.
Click to expand...
Meh. You know me, i want to see the numbers. Whats the contact resistance, whats the power being dissapated in the connector.

The only real world example i estimated numbers from in this thread is der8aur's unbalanced video, and current balancing wouldnt have solved his problem as far as i can tell.
 
Tsumi said:
Funny thing is the 8 pin melting at the PSU side using a 12vhpwr adapter is 100% due to the lack of current balancing on the 12vhpwr side.
Click to expand...
The load balancing protects the entire power cable from the supply to the GPU. As I've been saying and backing with analysis and evidence, this is most likely a load balancing issue and not a connector problem. It just shows at connectors first because they are most likely the weak link in the power delivery chain.

The other irony here is the pin that first burns up corresponds to the 12V link that was the best being the one with the lowest loop resistance back to the power supply. It gets overloaded, burns up, its resistance goes up and the fault moves to the next best and so on. The failure rate also accelerates as the link effectively goes from 6 to 5 to 4 to 3 to 2 to one power wire with the card black screening and smelling of burnt plastic somewhere down that countdown.

These cards are also an argument away from modular power supplies. That "extra" connector at the power supply is a risk. At the same time, with the 12VHWR end's propensity to fail, having modular cable there helps when one fails making it easier to replace. What an engineering cluster F.

If these were "normal" times, with 6000 series available at reasonable cost ($800) and performance improvement with properly designed power connectors/load balancing, cards like the 4090 and 5090 would likely be on people's "cards to avoid buying used" list, lumped in with the likes of the 970s and their gimped 4GB memory.
 
serpretetsky said:
The only real world example i estimated numbers from in this thread is der8aur's unbalanced video, and current balancing wouldnt have solved his problem as far as i can tell.
Click to expand...
I don't really want to start the argument again, but I really disagree with your statement.

der8aur's video and data are great, but one big limitation in it is he started collecting data after the failure started to cascade. This limits how far one can extrapolate the data. Ideally he would have started with all brand new components and measured overloads on pins before they became severe or before they significantly altered the resistances of the contact points. Any analysis starting with his data is compromised because the non-load balanced 5090 has already abused the connector and wiring for quite some time.

All load balancing has to do to avoid the problem is keep the individual currents of the power feeds lower than the max current allowed for each feed. It does not have to make them equal. If the individual currents are below the max spec for the connector and the connector actually meets its specs, the connector lasts forever even if the currents are not equal across its pins.

Load balancing design on these cards is an excellent example of compromise in engineering as it is an example of problem than cannot be solved economically using a perfection approach. Stepping back from perfection and figuring out "what can I live with" works really well here. Expecting all the conductors and contacts in the power cable to have exactly the same resistance is an example of the "perfection approach". Load balancing allows one to live with imperfect connectors and wiring that are available for reasonable cost, i.e. the compromise approach. Recall your "but it wastes more power" argument -- yeah, that's part of the compromise.
 
TXE36 said:
I don't really want to start the argument again, but I really disagree with your statement.

der8aur's video and data are great, but one big limitation in it is he started collecting data after the failure started to cascade. This limits how far one can extrapolate the data. Ideally he would have started with all brand new components and measured overloads on pins before they became severe or before they significantly altered the resistances of the contact points. Any analysis starting with his data is compromised because the non-load balanced 5090 has already abused the connector and wiring for quite some time.

All load balancing has to do to avoid the problem is keep the individual currents of the power feeds lower than the max current allowed for each feed. It does not have to make them equal. If the individual currents are below the max spec for the connector and the connector actually meets its specs, the connector lasts forever even if the currents are not equal across its pins.

Load balancing design on these cards is an excellent example of compromise in engineering as it is an example of problem than cannot be solved economically using a perfection approach. Stepping back from perfection and figuring out "what can I live with" works really well here. Expecting all the conductors and contacts in the power cable to have exactly the same resistance is an example of the "perfection approach". Load balancing allows one to live with imperfect connectors and wiring that are available for reasonable cost, i.e. the compromise approach. Recall your "but it wastes more power" argument -- yeah, that's part of the compromise.
Click to expand...
Current is meaningless in connectors and wires without knowing contact resistance or power dissipation. I already went over this. You can respond to that post if you disagree.
 
Feels like a lot of EE guys nerding out in this thread.
Forgive me electrical engineers and physics geeks, but I believe this long running thread can be summed up by "the 16-pin cables used in the 5090 (and higher NVIDIA GPUs) causes failures and needs to be replaced."
/end thread
 
sleepeeg3 said:
Feels like a lot of EE guys nerding out in this thread.
Forgive me electrical engineers and physics geeks, but I believe this long running thread can be summed up by "the 16-pin cables used in the 5090 (and higher NVIDIA GPUs) causes failures and needs to be replaced."
/end thread
Click to expand...

This x 1000000000000000000000000000000000000000000000000000000000000000000000000000
 
sleepeeg3 said:
Feels like a lot of EE guys nerding out in this thread.
Forgive me electrical engineers and physics geeks, but I believe this long running thread can be summed up by "the 16-pin cables used in the 5090 (and higher NVIDIA GPUs) causes failures and needs to be replaced."
/end thread
Click to expand...
I agree with your summation. Others here do not
 
serpretetsky said:
Meh. You know me, i want to see the numbers. Whats the contact resistance, whats the power being dissapated in the connector.

The only real world example i estimated numbers from in this thread is der8aur's unbalanced video, and current balancing wouldnt have solved his problem as far as i can tell.
Click to expand...
Doesn't matter what the numbers are. Load balanced would have caused instability and an investigation long before the connector got hot enough to melt if pin resistance was a problem. This also shows that simply adding headroom isn't enough- this connector was at 50% capacity and still melted because there was the possibility of all 300 watts going down one wire.
 
Tsumi said:
Doesn't matter what the numbers are. Load balanced would have caused instability and an investigation long before the connector got hot enough to melt if pin resistance was a problem. This also shows that simply adding headroom isn't enough- this connector was at 50% capacity and still melted because there was the possibility of all 300 watts going down one wire.
Click to expand...
Which doesnt fix the problem. Melted connector -> rma. Crashing gpu -> rma. Whats the improvement?
 
serpretetsky said:
Which doesnt fix the problem. Melted connector -> rma. Crashing gpu -> rma. Whats the improvement?
Click to expand...
Crashing GPU -> no physical damage -> try different cable or power supply first. That's the improvement. It's not the pins on the headers of the GPU/PSU that get damaged- it's the cable pins making poor contact.
 
Tsumi said:
Crashing GPU -> no physical damage -> try different cable or power supply first. That's the improvement. It's not the pins on the headers of the GPU/PSU that get damaged- it's the cable pins making poor contact.
Click to expand...
yeah you have a point. even better gets 2x 12vpwr connectors in parrallel, approach an acceptable failure rate, and forget all of this nonsense. i have mentioned before, but when i worked on backplanes/midplanes/motherboards i dont recall us doing any current balancing. just made sure we had over-specced connectors.
 
sleepeeg3 said:
Feels like a lot of EE guys nerding out in this thread.
Forgive me electrical engineers and physics geeks, but I believe this long running thread can be summed up by "the 16-pin cables used in the 5090 (and higher NVIDIA GPUs) causes failures and needs to be replaced."
/end thread
Click to expand...
If "nerding out" means actually understanding the problem, then why shouldn't we nerd away?

Let's say the contacts used for the 12VHPWR connector are rated for 10A each and there are six of them. 600W means 600/12 = 50A needs to be delivered. With 6 contacts, that is 50/6 = 8.33A per contact and we have a margin of 1.67A per contact.

However, the design of the video card does not account for the tolerance covering the resistance range of each of the contacts and winds up drawing 21A from a single contact (the der8baur example).

Did the video card fail or the connector?

I would say that is a video card fault as it overloaded the connector. Perhaps I'm being a bit pedantic here, but I think there is a chance that quad 8 pin PCIe connectors would have also burned up on the 5090 due to the lack of load balancing.

Other than the troll, there isn't anybody here arguing this problem does not have to be addressed. A robust solution requires digging down into the details and not reacting only to the symptom of a "bad cable" or "bad connector".
 
erek said:
do we understand the problem for sure tho? or what
Click to expand...
Given the available data, I'm comfortable enough with my understanding of it and I blame the video card design primarily. I've said before this situation is unlikely to end in a happy ending, although between WireView and the solder on adapters, it may work itself out. I highly doubt Nvidia is going to help in any way.

Also, as I've stated before, I don't own one of these cards so it's not really my problem beyond intellectual curiosity and I don't buy kilobuck video cards. The info learned here may become useful if 5090s ever get down to sub $500 used.

Jumped into the thread because enshittified engineering really peeves me, especially with very high dollar items like these.
 
erek said:
---
who
Click to expand...
Lgm The only person that keeps arguing with everybody that the connector is fine when you keep showing it's not.
erek said:
do we understand the problem for sure tho? or what
Click to expand...
Yes we do to much current is flowing though a pin till it melts. How to fix it is up to debate they needed higher safety factor and probably wouldn't hurt to have the power phase being broken up like they were on the 3 series and before. But honestly had they made the plug 25%+ larger contact area it probably would be ok the pins just aren't capable of carrying too much more than 10 or 11 amps before they start degrading so if it would had been an 8 power pins or just larger pins it probably would be ok.
 
tunatime said:
Yes we do to much current is flowing though a pin till it melts. How to fix it is up to debate they needed higher safety factor and probably wouldn't hurt to have the power phase being broken up like they were on the 3 series and before. But honestly had they made the plug 25%+ larger contact area it probably would be ok the pins just aren't capable of carrying too much more than 10 or 11 amps before they start degrading so if it would had been an 8 power pins or just larger pins it probably would be ok.
Click to expand...
I do think they likely would have been fine doing everything the same way except using a better plug or 4x the 8 pin PCIe connectors. Getting rid of the load balancing and going the 12VHPWR was likely a bridge to far.

That said, I believe one of the motivations to going to a smaller power connector was to shrink the board to enable them to use a larger heatsink. This challenge driven by 600W in an ATX case. Like it or not, these boards can be viewed as power pigs and remind me of the GTX480 and GTX580. Generations after those got more performance with less power. The 4090 and 5090 pretty much get more performance by more power.
 
Tick-tock...

Also from that pic on post 1270, is it always the +12V pins that are burned up? All things being equal, the ground returns should get burned too. Looks like there is enough additional ground path back through the PCIe connector and even the chassis to keep these from burning.
 
TXE36 said:
Tick-tock...

Also from that pic on post 1270, is it always the +12V pins that are burned up? All things being equal, the ground returns should get burned too. Looks like there is enough additional ground path back through the PCIe connector and even the chassis to keep these from burning.
Click to expand...
I think i might have seen one ground one burned up. But by far most of the photos in this thread are the +12V burning up.

edit: are you wondering how to balance current on the ground paths?
 
Last edited:
You must log in or register to reply here.
Back
Top