RTX 5090 FE Molten 12VHPWR

TXE36 · Saturday at 2:10 PM

Tsumi said:
In any situation where the combined design would overload pins, the separated current would generate significantly more heat in the marginal pins. What separated does prevent is runaway degradation from the positive feedback loop of increased current breaking down the contacts and causing more resistance, but it won't "last forever" as you claim.

The above is oversimplified to be useful in this discussion because all you are looking at is the connector, see below.

Realize that the engineering design of this power system is very advanced, it's not something a typical freshly minted 4 year EE degree will understand well, and frankly, even most grad school graduates would be challenged by it. This is something that takes real world experience to learn well and the true experts are highly specialized in switching power system design and that's all they do. I've been a practicing electrical engineer for 40 years and done some switching power design in the last 15 years. I find them fascinating devices, but there is a lot to them. When we have a big job, we farm it out to the guys who do it all the time. It's real easy to get too far in front of your skis with this topic.

Tsumi said:
Now, if we actually did want to do more engineering analysis, we'll need an analysis of how much heat the pins are capable of dissipating. That's going to be a complicated analysis that is dependent on the PCB temp, ambient air temp, and wire gauges in use. I checked and the spec calls for a maximum contact resistance of 5 milliohms, or 0.005 ohms, so technically the heat generated numbers would be half of what I calculated above. That does imply the design has very little tolerance for heat in the connector.

That looks like a really hard way to approach what is a simple problem.

The simple approach is to stand on the shoulders of the connector manufacture and use their max current and ambient temp ratings and design the system to stay below them. Contact resistance comes in to bound the max current in each wire to the max current for each pin. This would actually guide one to the number of shunts required.

Think of the problem from the other direction, if the current in each pin does not exceed the spec max, and the ambient temperature is within specs, why would the connector burn up? I would argue, if that occurred, that the connector did not meet its datasheet spec, which is still possible here.

That said, how are you accounting for the variation in wire resistances from the power supply? If that variation is large enough, it will channel most of the load current into one or two of the wires then into one or two contacts and the connector is overloaded and burns up. Current balancing will prevent this because it will force all or nearly all the wires to carry the load thus keeping all the contacts of the connector in spec and the connector lasts forever.

People coming to this thread to seek reassurance that their 5090 won't burn up are going to be disappointed as that's not happening short of things like that PCIe 8 adapter posted earlier. At this point, there is no way to guarantee the stock boards are "fine".

TXE36 · Saturday at 3:05 PM

Let's try a different angle. The power connector has six 12V pins. 600W@12V is 50A. Nominally divided across six pins, that's 8.33A each. What if most of the current only uses two pins because the total resistance of those two paths is a lot lower than the other 4? If all the 50A goes through those two pins, that's 25A each. Earlier I posted an IR picture from der8auer that showed nearly exactly this. The caveat is we don't know for sure how much the currents are in the two wires compared to the other four, but the picture shows they are very different:

I would argue the current imbalance is the cause and the connector failure is the result.

What would current balancing do? Simple. The current distribution above is controlled solely by the path resistances. If the power supply forced the current to flow in three two wire pairs, the worst case scenario is you would have 3 wires carrying the load -> 50/3 = 16.67A. If you used six shunts, you could force the current in each wire to be about 8.33A practically independent of their resistances (the resistances have to be within reason).

However, der8auer did us a huge favor: He actually measured the currents (with a clamp on meter I suggested earlier here) in the wires that correspond to that picture: der8auer measured wire currents

23A
2A
5A
11A
8A
3A

Look at the IR picture again, one of the hot wires is hotter than the other. This one is 23A. The next hottest is the 11A one. The total load is 52A (624W) the hot pin is shouldering 276W and the other hot is shouldering 132W. The other 4 slackers are shouldering 216W between them.

This is it, full stop. der8auer did a lot of work and has the evidence to back it up. 23A through a connector contact that nominally should be 8.33A is nearly a 3x overage. Even a "conservative" design which used a 2x rated contact (16.67A) would have been overloaded by 1.4x.

I wish I had looked into his video more deeply earlier. He demoed the failure in real time and collected the data. There are no statistics that can defeat this.

serpretetsky · Saturday at 3:19 PM

TXE36 said:
However, der8auer did us a huge favor: He actually measured the currents (with a clamp on meter I suggested earlier here) in the wires that correspond to that picture: der8auer measured wire currents

23A
2A
5A
11A
8A
3A

edit: let me know if something here doesn't make sense.

edit2: Fixed some of the current values predicted by Non-balanced calculation that didn't matchup with original assumption. Thanks TXE36. (I had copied a resistance value incorrectly)

TXE36 · Saturday at 3:48 PM

serpretetsky said:
View attachment 831724
View attachment 831725

edit: let me know if something here doesn't make sense.

Not 100% sure how you are doing the calcs, but they look close. I don't know why the 8A line is only 6.94A in your non-balanced case.

You realize your confirming the argument I've made all along? I said I don't care about the wasted power, I said I wanted the connector to be within spec. Yeah, I'm "wasting ~25W", but l@@K at the balanced connector current:

I think a limitation on the math is we are assuming 12V all the time -- that's probably not fully accurate. Ideally you would want to measure the voltage at each pin too.

One way an individual can get a warm fuzzy on this is to monitor the current in the six wires, either with a clamp on ammeter or some gizmo you can plug in line with the connector that will tell you the currents and monitor the balance. At least this stops the needless disconnect and connect cycles. The problem with the clamp on meter is you have to unbundle the harness and you have to remember to do it.

A device that would monitor the balance and automatically cut the power if too large a current imbalance happened would save the card/connector from massive damage, but the wiring harness and the connector on the video board would still have be replaced. This solution can catch the problem, but not solve it (current balance on the PCB most likely would solve it).

If I had one of these cards, I'd really look at that 4x PCIe 8 adapter. Perhaps compromise with monitoring and go the adapter route if the 12VHPWR connector fails.

TXE36 · Saturday at 4:00 PM

erek said:
Hope

I think we are seeing a bit of that today

.

serpretetsky · Saturday at 4:03 PM

TXE36 said:
Not 100% sure how you are doing the calcs, but they look close. I don't know why the 8A line is only 6.94A in your non-balanced case.

You realize your confirming the argument I've made all along? I said I don't care about the wasted power, I said I wanted the connector to be within spec. Yeah, I'm "wasting ~25W", but l@@K at the balanced connector current:

Good catch. Probably artifact of me trying to approximate and deduce. You are correct assuming 12V could also slightly throw off the values. I'm not going to comment whether or not im confirming your argument. Make your own decision. 25W vs 51W waste heat. You decide if that's important.

(you already know what I think)

Tsumi · Saturday at 4:08 PM

I guess I can see where you're coming at. If all the connectors are technically within spec at less than 0.005 ohms but most are around 0.003, one is at 0.005, and one is at 0.001, it will create the kind of current imbalance that will make the 0.001 ohm pin overheat and fail. Each pin will have to be tested and binned to provide equal resistance regardless of whether or not they're in spec in a combined bus scenario. Separated pins are guaranteed to not have a current imbalance even for pins that are in spec but not closely matching in resistance. The total heat generated doesn't matter- it's the concentration of current and heat in one pin.

TXE36 · Saturday at 4:14 PM

serpretetsky said:
Good catch. Probably artifact of me trying to approximate and deduce. You are correct assuming 12V could also slightly throw off the values. I'm not going to comment whether or not im confirming your argument. Make your own decision. 25W vs 51W waste heat. You decide if that's important.

(you already know what I think)

You don't care about trying to shove 24A though a contact rated for 8.33A? Your own spreadsheet confirms what I've been saying, I circled the relevant parts in the red and green boxes.

TXE36 · Saturday at 4:20 PM

Tsumi said:
I guess I can see where you're coming at. If all the connectors are technically within spec at less than 0.005 ohms but most are around 0.003, one is at 0.005, and one is at 0.001, it will create the kind of current imbalance that will make the 0.001 ohm pin overheat and fail. Each pin will have to be tested and binned to provide equal resistance regardless of whether or not they're in spec in a combined bus scenario. Separated pins are guaranteed to not have a current imbalance even for pins that are in spec but not closely matching in resistance. The total heat generated doesn't matter- it's the concentration of current and heat in one pin.

Go beyond the connector and look at the resistance from the GPU input all the way back to the ATX power supply output. Again, it's not just the connector.

If one of my 12V lines is twice as long as another, it will have 2x the resistance. Not likely, but same concept. How well matched are the lengths of the power wires?

Look at der8auer's video and measurements - he is showing it in real time.

Tsumi said:
The total heat generated doesn't matter- it's the concentration of current and heat in one pin.

That's it!

Well, technically one wire too. The weak point is the connector, but the wires are likely being driven over spec.

serpretetsky · Saturday at 4:25 PM

TXE36 said:
You don't care about trying to shove 24A though a contact rated for 8.33A? Your own spreadsheet confirms what I've been saying, I circled the relevant parts in the red and green boxes.

Its certainly a concern. But here is what you aren't looking at. Look at pin2 estimated resistance. You want to shove 8amps through that? Have fun. This connector is already out of spec from the get-go. You're trying to apply some sort of current balancing bandaid onto a leg that is ripped off.

TXE36 · Saturday at 4:45 PM

serpretetsky said:
Look at pin2 estimated resistance. You want to shove 8amps through that? Have fun.

Why not? What specification does it violate? All I'm looking at is that nice 9A on each connector pin.

Keep in mind, for this current balancing scheme to work the range of resistances and spread of resistances all the way back to the ATX 12V power supply have to be reasonable, this is real world, not an academic math problem. What is the resistance of the wire itself compared with the contact resistance?

You are so close to understanding this. Extend your thinking beyond the connector. I think going backwards to the power supply is easier and der8auer is showing that clearly. Think, how could a "high resistance contact" make two of those six wires glow? Think of the contact resistance + the wire resistance back to the 12V power supply. ETA: It's not just the connector pin getting hot, it is the whole bloody wire!

der8auer is showing what the system is doing - come up with your own explanation for it. Note he almost never talks about the connector, and Buildzoid doesn't either. There is no light between what der8auer presented and Buildzoid - they are precisely aligned on the topic and have this issue nailed.

serpretetsky said:
This connector is already out of spec from the get-go.

How so?

serpretetsky · Saturday at 4:53 PM

TXE36 said:
Why not? What specification does it violate? All I'm looking at is that nice 9A on each connector pin.

Keep in mind, for this current balancing scheme to work the range of resistances and spread of resistances all the way back to the ATX 12V power supply have to be reasonable, this is real world, not an academic math problem. What is the resistance of the wire itself compared with the contact resistance?

You are so close to understanding this.

Connector resistance is extremely important for any system that transmits any reasonable amount of power. This true for all connectors. If the datasheet doesnt list it just means they ommited it. Ive done enough work in this thread. Find a molex microfit 3.0 datasheet. Find one with contact resistance specified. Post your results here. Drop the condescending teacher attitude. Its not helping you.

TXE36 · Saturday at 5:02 PM

serpretetsky said:
Connector resistance is extremely important for any system that transmits any reasonable amount of power. This true for all connectors. If the datasheet doesnt list it just means they ommited it. Ive done enough work in this thread. Find a molex microfit 3.0 datasheet. Find one with contact resistance specified. Post your results here. Drop the condescending teacher attitude. Its not helping you.

Connector resistance isn't the only thing and it appears you have tunnel vision about it.

Didn't mean to offend, don't have a dog in this hunt, and the results posted by der8auer and Buildzoid are 100% conclusive. I have no motivation for further research.

So much for the hope...moving on.

Tsumi · Saturday at 5:38 PM

serpretetsky said:
Connector resistance is extremely important for any system that transmits any reasonable amount of power. This true for all connectors. If the datasheet doesnt list it just means they ommited it. Ive done enough work in this thread. Find a molex microfit 3.0 datasheet. Find one with contact resistance specified. Post your results here. Drop the condescending teacher attitude. Its not helping you.

The thing you're not taking into consideration is that the pin connection resistance can be technically in spec yet still differ enough to cause a significant current imbalance. In this scenario, a combined bus can fail while the separated pins will continue working just fine.

The specification is less than 0.005 ohms and less than 0.006 ohms when a sideways force is applied to it. It doesn't say anything about binning to ensure pin resistance is the same across all pins in the connector.

AVATARAT · Saturday at 6:19 PM

TXE36 said:
Connector resistance isn't the only thing and it appears you have tunnel vision about it.

Didn't mean to offend, don't have a dog in this hunt, and the results posted by der8auer and Buildzoid are 100% conclusive. I have no motivation for further research.

So much for the hope...moving on.

The main problem here is that the only thing we can do is monitor the connector's condition and replace it... Simply because we can't balance the PCB—right?

TXE36 · Saturday at 6:42 PM

AVATARAT said:
The main problem here is that the only thing we can do is monitor the connector's condition and replace it... Simply because we can't balance the PCB—right?

It sounded like the guy(s) doing that quad PCIe 8 adapter were talking about putting some sort of shunting in there, that should help but may be tricky as there is still only a single 12V feed (shunt) into the card.

The reason this works better on the PCB is you have a multi-channel step down regulator there that changes the 12V to whatever the core voltage is. Traditionally this has had 3 12V inputs (3 shunts) and the regulator controller can pick how much current to load down each input by monitoring the shunt voltages. While there are three input channels, there are lots more output channels on the regulator to spread all of this out so the MOSFETs don't blow up. The board has to have this 12V to core regulator, the losses from it are already baked in.

To do this before the single 12V input to the card, you would create three 12V to 12V channels each with a shunt and regulate them. There is such a thing as a 12V to 12V switching regulator, and it works the same way as other switching regulators. However, all regulators have loss, so this 12V to 12V regulator would have to be rated for 600W -- sounds pretty big and how do you heatsink it? An external 12V to 12V regulator would introduce a lot more waste heat that is not already baked in.

Recall that another way to treat this would be with a better connector, 4 PCIe 8 adapters soldered to the 12VHPWR pads may be good enough. Their early result with running 800W through it looks promising. If that wasn't quite enough, use 8 PCIe 8 adapters - when the graphics card is worth north of $10K, ridiculous ideas like this sound feasible.

Now here is a really off the wall idea, enabled by the value of these things: Take the BIOS, GPU and memory (the expensive stuff) and put it on a new PCB with a proper power input section. The guy at Northwest removes and installs GPUs and memory all the time (making it look easy - that guy has mad skills). Would not be cheap, and in normal times with normal stuff not likely economical, but in our nutty times I dunno...

serpretetsky · Saturday at 8:07 PM

Tsumi said:
The thing you're not taking into consideration is that the pin connection resistance can be technically in spec yet still differ enough to cause a significant current imbalance. In this scenario, a combined bus can fail while the separated pins will continue working just fine.

Yeah you might be right. I would have to think on it.

I mean youre definitely correct in that you could easily violate the max current spec on a pin that way. I have no disagreement there.

The part im not sure about is if that actually matters so much on pin/wire that has such little contact resistance (basically tha pin/wire is the superstar of the show, and can carry that extra load anyways). In other words the manufuctere could have, in theory, specced that individual pin/wire for much higher current and it would have been fine. But im not sure.

serpretetsky · Saturday at 8:25 PM

TXE36 said:
Didn't mean to offend,

Thats ok. Sorry that i was so short with you.

serpretetsky · Sunday at 12:02 PM

TXE36 said:
Connector resistance isn't the only thing and it appears you have tunnel vision about it.

Yes, it probably does appear that I have tunnel vision, and I have been doing a disservice to you and Tsumi for not clarifying why it might appear im so obsessed with waste power dissipation and electrical resistance.

So I'm going to say something that might be controversial:

There is no such thing as current carrying capacity for a metal connector or wire.
Instead the main limitation is the max temperature allowed by the materials used, and the acceptable voltage drop for downstream applications.
Insulation melts/burns at certain temperatures.
Metal contact points can degrade after repeated heating/reheating to high temperatures.
Degraded contacts have more electrical resistance.

Temperature is a function of the geometry of the material, the thermal conductivity, and the power injected into the material (this is the waste heat that I always talk about)
Perhaps I didn't clarify very well earlier, I dont actually care that power is wasted in the wire.
Like TXE36 mentioned, even 50 watts is kind of trivial when compared against 600 watts.

What I care about is waste power means heat. It is the source of all heating in our connnector.
5 Watts in a small connector insulated by plastic housing might be ok. 25 watts could be pushing it too far. it depends on the connector.

How do we determine waste power in the connector? I think you guys all know this one already, P=IV. Or for an ohmic system you can do P=I^2R. The R in that equation is why I also put emphasis on resistance in my calcs.

Here are some reads to backup what im talking about:

https://www.researchgate.net/public...with_Respect_to_the_Current_Carrying_Capacity

"Fig. 3 shows the first part of the systematic design. It is the procedure for determining the maximum permissible powerloss of a connector contact"

"The current carrying capacity of connectors is significantly influenced by the connector resistance"

"In order to dimension the maximum permissible connector resistance, the heat flow into the wire is used [21]. The maximum heat flow can be calculated, if the temperaturedifference at the head of the wire ∆ϑ__ is defined. Theheat flow depends on the permissible temperature difference atthe wire as well as the wire cross-section."

https://www.materion.com/en/insights/blog/connector-current-capacity

"Occasionally, a designer may ask how much current can safely be run through a given connector design. While this is a simple question, the answer is very complex. The maximum safe current depends on three parameters: connector geometry, electrical resistance, and maximum allowed temperature rise. "

"it is common to see heavy-duty extension cords rated at certain current levels, such as 10 or 15 Amps. This is the maximum recommended current that can be passed through the cord without compromising safety. This level is known as the cord’s current carrying capacity, or ampacity. This leads to a common misconception that a given material has a maximum amount of electrical current it can handle without failure. This is false, since there are many different possible failure mechanisms for a given current-carrying situation. "

https://www.weipuplug.com/knowledge...ing-affects-system-reliability-and-lifecycle/

"A connector current rating is not just a number printed on a plug body. It is connected to contact size, contact pressure, contact resistance, heat rise, cable size, ambient temperature, enclosure conditions, mating quality, and how often the connector is used. When these factors are ignored, a connector may overheat, age faster, loosen under vibration, increase voltage drop, or become a hidden failure point in the system."

"Current rating is closely linked to temperature rise. Connector engineering references explain that rated current is determined by the temperature rise created when current flows through the connector, and safe current capacity depends on connector geometry, resistance, and allowed temperature rise."

https://connectorselection.com/articles/connector-current-voltage-ratings-explained/

"Manufacturers evaluate current limits based on a specified temperature rise (ΔT), typically a 30°C increase above ambient room temperature. If a contact's temperature climbs beyond its maximum rated operating temperature (often 105°C to 125°C for standard plastics), the plastic housing will soften, lose structural integrity, and fail to hold the contacts in place."

This is why while everyone is obsessed with trying to balance current on a connector, my simple question to them is WHY??? You see that pin that is carrying most of the current? There is a reason that most of the current is going there in the first place. It has low electrical resistance.
Im not saying that a passive unbalanced connector is the most optimal way to send current down multiple wires, probably some form of power balancing would be more optimal(and more complicated). But current balancing is usually the WRONG thing to do.

Ok... so why do some datasheets not emphasize this? Well, what would they put? individual contact resistance on per-pin basis? They can't do that, they can just give you nominal/max values.

https://www.molex.com/content/dam/molex/molex-dot-com/en_us/pdf/datasheets/987650-5984.pdf?inline

"Contact Resistance: 10 milliohms"
is this nominal? is this max? I dont know. Tsumi mentioned somewhere he found that 12VHPWR is 5 milliohms max i beleive (please correct me)
Either way, this is the ballpark we should be working in. 100 millohms is most likely way too high, and you can know it once you try to simulate it and see how the power looks.

serpretetsky · Sunday at 12:08 PM

Tsumi said:
The thing you're not taking into consideration is that the pin connection resistance can be technically in spec yet still differ enough to cause a significant current imbalance. In this scenario, a combined bus can fail while the separated pins will continue working just fine.

The specification is less than 0.005 ohms and less than 0.006 ohms when a sideways force is applied to it. It doesn't say anything about binning to ensure pin resistance is the same across all pins in the connector.

So now that I posted my wall of text above. The question is: you have some wires that are well within spec, but one of them is actually way further below spec. How does balanced current vs unbalanced look like in this case:

First block of 4 is all resistances at 10millohms. Second block of 4 is now set one of them to 2mohm.s You're not wrong. The total waste power is still better with the non-balanced version. However, look at the max power of pin4 in the non-balanced version. Its slightly higher than any max pin power of the balanced version. Which one is better? Hard to tell, would actually need to test it in the real world to see which is hotter.

However, here's something to think about: that really hot one in the non-balanced version? What happens to resistance when metals and metal contacts heat up? It goes up. There is a chance that hot one could negative feedback and balance back out.

serpretetsky · Sunday at 12:19 PM

TXE36 said:
Look at the IR picture again, one of the hot wires is hotter than the other. This one is 23A. The next hottest is the 11A one. The total load is 52A (624W) the hot pin is shouldering 276W and the other hot is shouldering 132W. The other 4 slackers are shouldering 216W between them.

I have nothing techincally against this, i just dont like when people specify "the load" carried by each wire in watts. Its actually the downstream load, and it can be misleading since it has little to do with the wire.

Not 100% sure how you are doing the calcs, but they look close. I don't know why the 8A line is only 6.94A in your non-balanced case.

I corrected that 6.94A line, i copied a resistance value incorrectly from the initial spreadsheet into my python script. thanks.

For calculating balanced vs non balanced I have a python script that calculates both. My assumption is that the downstream VRM is a constant power load, and I can set it arbitrarily to whatever we want. For the non-balanced version it was actually just kirchoff's current law and then solving a quadratic equation. For the balanced version I had a brain fart and couldn't figure out an analytical way to do it, so instead the script guess/checks various currents until it roughly matches the downstream power consumption. It first finds the max power possible with the input resistances and downstream power assumptions, and uses that as an upper boundary, and then binary searches downward till i get enough precision. 1024 iterations seems to be enough (i guess float64 values have about 10bits positve and negative for the exponent, so 2**10 = 1024, kind of makes sense)

Think, how could a "high resistance contact" make two of those six wires glow? Think of the contact resistance + the wire resistance back to the 12V power supply. ETA: It's not just the connector pin getting hot, it is the whole bloody wire!

In my experience wires themselves rarely have a lot of variation in resistance, it's typically the contact or the crimp point.
I say this because the 2 glowing wires are the low resistance ones, and because the contact/crimp point has relatively low resistance the entire chain has a more even resistance distribution, hence the entire wire heats up.

TXE36 · Sunday at 2:50 PM

serpretetsky said:
I have nothing techincally against this, i just dont like when people specify "the load" carried by each wire in watts. Its actually the downstream load, and it can be misleading since it has little to do with the wire.

In my experience wires themselves rarely have a lot of variation in resistance, it's typically the contact or the crimp point.
I say this because the 2 glowing wires are the low resistance ones, and because the contact/crimp point has relatively low resistance the entire chain has a more even resistance distribution, hence the entire wire heats up.

Then just leave the load carried by each in amps. But realize the power dissipation in the wire will be those amps times the resistance of the wire. Power dissipation in the wire will raise its temperature.

Wire itself does tend to be uniform and there is much more variation at the contact and crimp. These days contacts can be very problematic as the cheapening of connectors means some manufactures cheat on pin plating. At work we have seen failures with old school 9 pin serial connectors because of this -- I've never seen that before.

I skimmed it, but your analysis of max currents above looks correct - yes power limits on wires are mainly determined by allowable temperature rise. Since connections have non-zero resistance, there will always be a temperature rise in them with current.

The bottom line is der8auer's work shows the current imbalance happening in real time. One may not fully understand the physics (I don't understand all of it), but it's right there in his data. It's irrefutable, so why is it happening? What could be done to address it? I think it is due to the variance in the resistances of the 24 connections and 12 wires (assuming a single power cable between the card and power supply). Yes, those variations are likely small, but it appears they are large enough to cause 23A to be steered into one of the power feeds. If not, then what's causing that?

That said, I would encourage you to take a step back and think about what you are doing and what you are trying to achieve. What are you trying to achieve and why?

serpretetsky · Sunday at 3:53 PM

TXE36 said:
Then just leave the load carried by each in amps. But realize the power dissipation in the wire will be those amps times the resistance of the wire. Power dissipation in the wire will raise its temperature.

agreed.

The bottom line is der8auer's work shows the current imbalance happening in real time. One may not fully understand the physics (I don't understand all of it), but it's right there in his data. It's irrefutable, so why is it happening? What could be done to address it? I think it is due to the variance in the resistances of the 24 connections and 12 wires (assuming a single power cable between the card and power supply).

Agreed that the current imbalance is from resistance variation. But the next step is where I and a lot of others differ. Other's argument at this point is "let's balance the current". This is a fundamental misunderstanding of what is actually the problem and if there even is a problem. The current imbalance is a symptom. The actual question is "what is the temperature of the individual pins". This is hard to simulate, i dont know how to do it. But what I CAN do is compare power dissipation from examples that we assume are good, or examples that we assume are bad, and see how it compares. That is why I focus on comparing power dissipation. This is why I focus on pin resistance.

Let's say the power dissipation actually is too much. I mean, those 2 wires from der8auer's video looks pretty hot. So again, others will say "let's balance the current". But balancing the current in many situations is actually worse. More power is being dissipated in the connector. That means the average temperature of the connector will be even higher. Perhaps the per-pin temperature will be lower? In the case where you have 1 bad pin, it is always worse to current balance. Both per-pin and total power dissipated in the connector is higher. If there are more than 1 bad pins? Current balancing total power is still worse, however per-pin is sometimes better. It's not clear to what's better.

Furthermore, when you do current balancing, you are putting the stress back onto the bad connections. If there is any sort of fatigue or positive feedback, then these pins will get worse and worse until failure. If you leave most of the current on the good pins, there is actually some chance for negative feedback to help equalize things a little.

There could be situations where current balancing could objectively help. For example, I noticed in gamers nexus video that they found situations where the solder joint between a PCB and the connector is poor. There specific example was an adapter that connected to the GPU I believe, but let's, for example, pretend the gpu itself actually has a poor solder joint. It could make sense to force more current onto that pin, even if it has poor resistance, simply because a solder joint on a PCB can withstand much higher temperatures than a pin in a plastic housing. Like you said, who cares about wasted power, as long as the heat is dumped somewhere where it can disappate quickly, and the materials can withstand it.

That said, I would encourage you to take a step back and think about what you are doing and what you are trying to achieve. What are you trying to achieve and why?

I want people to take a step back and think about what they are claiming is the problem, and what they are claiming is the solution.

TXE36 · Sunday at 4:16 PM

How do you propose to solve the hot wire problem?

You may not like the current balancing solution, but it has worked in the past on Nvidia video cards. I have thought about what I'm claiming is the problem and I've backed up every claim I've made. Several claims by others have been backed up as well.

About the only interesting thing left here that has not been explored is 3090 TIs with the new connector that have burned up in unmodified form. Have any been found? They have the same type of current balancing as earlier Nvidia cards.

Edit: NM, you answered that above. ~~Why are those two wires hot in der8auer's video and current measurements? You certainly can't fault current balancing for that.~~

TXE36 · Sunday at 4:38 PM

How about yet another tack?

Think of each power wire individually. They are all rated at 8.33A. If you don't want them to burn up, you would fuse them at 8.33A. Since blowing a fuse is inconvenient, we'll let the power supply monitor the current in each of the six legs and if any of them exceed 8.33A, the power supply clicks off. Assuming all the wiring and connectors can handle 8.33A, nothing will burn up.

How do you get a 5090 to work with this power system? At 23A in one leg, the components in der8auer's example would shut down that power supply immediately.

And before you say my example is unreasonable, it's not far from what good commercial power supplies do. I'll take the supplies I use as an example, a Delta DPS-800LB A that came out of a high end ($5000 MSRP) HP workstation:

If you look, the 12V portion of the supply is rated for 64A (768W). There are several 12V outputs with one being rated at 18.6A and the rest at 18A. If I try to pull more than that 18A or 18.6A from those power rails, the power supply will click off. I know it as I've done it.

Also note that this supply likely would not be very good for a 5090 as getting 50A of capacity would practically require consuming 3 of the 18A rails just for the video card. Even if you work that out, it only leaves 168W for the rest of the 12V loads and then the whole thing has to stay under 800W.

However, it will drive a pair of 1080TIs or Titan XPs with a 5800X3D quite well and it is well worth the ~$30 I paid for it.

serpretetsky · Sunday at 4:39 PM

TXE36 said:
You may not like the current balancing solution, but it has worked in the past on Nvidia video cards.

Sorry to repeat myself, but how do we know it has worked? You say "look at the 1080ti, it was fine". I say "yes, look at the 1080ti, it had GOOD CONNECTORS, regardless of current balancing"

About the only interesting thing left here that has not been explored is 3090 TIs with the new connector that have burned up in unmodified form. Have any been found? They have the same type of current balancing as earlier Nvidia cards.

It's a good question. There are unfortunately, more variables with the 3090ti that make it hard to compare against 4090/5090. 3090ti has 450 watt power consumption, quite a bit less than 600 watt. Second, like Tsumi pointed out, when you split up the 12V nets like 3090ti and previous cards did, it means each VRM can easily monitor its own 2 or 3 pins (however many, i forget). This is nice, because if there is a problem, the card should, in theory, just shutdown due to undervoltage on the VRM input (you cant do this on an unbalanced card without first introducing some shunts to isolate out the pins).

However, this is not nice, because you still have a problem, but instead of melting connector, you have a card that intermittenly shuts down.

Now let's think about this from a user perspective. You have a 3090ti that intermittenly shuts down, nothing interesting is happening. Do you post pictures and reddit posts claiming "3090ti connector failure". Or do you silently just send the card in for an RMA, and tell people it was just an unstable card, and they sent you a new one and you forgot the whole mess.

edit: or send the psu/ adapter in for rma

serpretetsky · Sunday at 4:45 PM

TXE36 said:
How about yet another tack?

Think of each power wire individually. They are all rated at 8.33A. If you don't want them to burn up, you would fuse them at 8.33A. Since blowing a fuse is inconvenient, we'll let the power supply monitor the current in each of the six legs and if any of them exceed 8.33A, the power supply clicks off. Assuming all the wiring and connectors can handle 8.33A, nothing will burn up.

How do you get a 5090 to work with this power system? At 23A in one leg, the components in der8auer's example would shut down that power supply immediately.

Yes, this is pretty similar to having separate pins for separate vrms like nvidia's older GPUs had. This works. To re emphasize, it does have the benefit that it's easier to spot a problem and shutdown before the connector actually melts. I'm just arguing that isn't a complete solution. Now instead of connector burning, you have PC that randomly shuts off.

AVATARAT · Sunday at 5:00 PM

serpretetsky said:
Let's say the power dissipation actually is too much. I mean, those 2 wires from der8auer's video looks pretty hot. So again, others will say "let's balance the current". But balancing the current in many situations is actually worse. More power is being dissipated in the connector. That means the average temperature of the connector will be even higher. Perhaps the per-pin temperature will be lower? In the case where you have 1 bad pin, it is always worse to current balance. Both per-pin and total power dissipated in the connector is higher. If there are more than 1 bad pins? Current balancing total power is still worse, however per-pin is sometimes better. It's not clear to what's better./home/nikolayd/Desktop/

Here's some information—my cable is about 10 months old.
Here are two tests at different power levels to show you the temperatures:
~26C abient

TXE36 · Sunday at 5:25 PM

serpretetsky said:
Sorry to repeat myself, but how do we know it has worked? You say "look at the 1080ti, it was fine". I say "yes, look at the 1080ti, it had GOOD CONNECTORS, regardless of current balancing"

How do you know those connectors, by themselves are good enough? The littlegreenman posted several pictures of burned up PCIe 8 connectors. All the 1080TIs had current balancing and I suspect the practice has gone back a long way. I think all the 3090 TIs had current balancing.

serpretetsky said:
It's a good question. There are unfortunately, more variables with the 3090ti that make it hard to compare against 4090/5090. 3090ti has 450 watt power consumption, quite a bit less than 600 watt. Second, like Tsumi pointed out, when you split up the 12V nets like 3090ti and previous cards did, it means each VRM can easily monitor its own 2 or 3 pins (however many, i forget). This is nice, because if there is a problem, the card should, in theory, just shutdown due to undervoltage on the VRM input (you cant do this on an unbalanced card without first introducing some shunts to isolate out the pins).

The 3090 TI is the *perfect* card to use as a comparison because it has only one thing in common with the 5090 - that 12VHPWR connector. It does not matter that its TDP is 450W as there was a 5080 reported upthread that failed at 300W. Therefore, we have reports of 12VHPWR connectors failing at 300W. Do we have any reports of 3090TIs burning connectors at 300-450W?

In normal operation, the VRM on the 3090 TI pulls power as appropriate from each of its three power feeds. In normal operation, it does not overload any of them. If a fault developed in any one of those legs that prevented it from supplying enough power output, the card would shut down. The card is designed to supply 450W to the board from the three feeds. They have already engineered how much current would be required from each feed to work normally plus some tolerance. The card itself has the ability to both sense faults and keep the connector currents within reason. The 5090 does *not* have this capability. I'll give an example.

Like the 5090, the 3090 TI has six power feeds. Because of the current splitting, the feeds are divided into 3 pairs. Each pair is rated for 50A/3 = 16.67A, however the card is limited to 450W, so that becomes 450W/12V = 37.5A and thus 12.5A for each pair. Nominally it is 6.25A per wire. While operating, the core regulator monitors the currents in all three shunts while it is also supplying voltages to the card. If any of those three currents exceeds 12.5A it can dial back the load power or shut off the core as it has detected a fault condition. Worst case scenario is 12.5A on a pair with all the current in one wire -- this is half of what the 5090 has demonstrated.

The 5090 has one power feed and is rated for 600W. Its voltage regulator will not detect a fault until the single shunt current is 50A, this is 4 times higher than the 3090 TI. der8auer's 5090 is plodding along pretty as you please oblivious to the 23A fault in one of its feeds. A fault the 5090 caused because the resistivity imbalance of the power pairs cause it to draw too much current through one of them, something which until we find a 3090 TI with a burned 12VHPWR is unlikely to happen with a 3090 TI. Recall we have already seen these connectors fail without current balancing at 300W.

erek · Sunday at 5:29 PM

5090 fe
melted 12vhpwr

https://x.com/unikoshardware/status/2111895726729097584

TXE36 · Sunday at 5:36 PM

More to toss in the mystery. It looks like the 5080 has two shunts and is more like the 4090 which Buildzoid reported also has two, I initially assumed it also had one like the 5090. This suggest that two shunts are not enough as 5080s and 4090s have burned these connectors.

Makes the quest to find a burned 3090 TI with one of these connectors even more interesting. Were 3090 TIs ever made with PCIe 8 pin connectors and not the 12VHPWR?

Here is the 5090 with the one shut circled in red:

Here is the 5080:

I don't have access to the cards, the pics above came from Techpowerup. In the 5080 pic, there are also two more low value resistors below the LR22 below the red boxes. Hard to tell what these are, maybe shunts for the memory power supply?

serpretetsky · Sunday at 5:37 PM

TXE36 said:
How do you know those connectors, by themselves are good enough? The littlegreenman posted several pictures of burned up PCIe 8 connectors. All the 1080TIs had current balancing and I suspect the practice has gone back a long way. I think all the 3090 TIs had current balancing.

I dont know they're good enough. That was just my counter argument. Im pointing out that 2 variables have changed, and you are claiming that it's the current balancing that made the 1080ti reliable. How did you confirm that it isn't the beefier connector?

The 3090 TI is the *perfect* card to use as a comparison because it has only one thing in common with the 5090 - that 12VHPWR connector. It does not matter that its TDP is 450W as there was a 5080 reported upthread that failed at 300W. Therefore, we have reports of 12VHPWR connectors failing at 300W. Do we have any reports of 3090TIs burning connectors at 300-450W?

I suppose this is where we will simply disagree, because this is where I'm going to bring samples and failure rates. A single 5080 that failed at 300W doesn't tell me anything. I mean... it tells me that single sample was so utterly broken from the factory that it failed even at 300W. I dont know what was wrong with that single card. Was the 12VHPWR connector defective from molex or wherever they get the connector from? Bad solder job? Single point failures are not necessarily design failures. I can't extract anything meaningful from a single failure.

edit: btw, im going to logoff of hardforum for a while. Im kind of exhausted

Edit2: this was not a trivial amount of work for me

TXE36 · Sunday at 5:50 PM

serpretetsky said:
I dont know they're good enough. That was just my counter argument. Im pointing out that 2 variables have changed, and you are claiming that it's the current balancing that made the 1080ti reliable. How did you confirm that it isn't the beefier connector?

edit: btw, im going to logoff of hardforum for a while. Im kind of exhausted.

I don't know for sure if is the beefier PCIe 8 connector and don't believe I made that claim. In fact, when commenting on that quad PCIe 8 adapter, I mentioned that I wasn't sure it was enough.

Me too. As you can tell I like discussing this stuff and it has been a nice detraction from my usual Sunday dreads. Got 4 more Mondays starting tomorrow before another Friday gets here: M-M-M-M-F

serpretetsky said:
I suppose this is where we will simply disagree, because this is where I'm going to bring samples and failure rates. A single 5080 that failed at 300W doesn't tell me anything. I mean... it tells me that single sample was so utterly broken from the factory that it failed even at 300W. I dont know what was wrong with that single card. Was the 12VHPWR connector defective from molex or wherever they get the connector from? Bad solder job? Single point failures are not necessarily design failures. I can't extract anything meaningful from a single failure.

That single failure is pretty hideous that it happened at 1/2 the connector's rated power. That's not a near miss, it's a face plant.

For me it is real simple, the previous cards with three shunts didn't have this problem. It is very logical to ask could the lack of shunts on the failing boards be the root cause? Again, the 3090TI may demonstrate a case where the 12VHPWR connector is fine and dandy -- if that's the case, it's more evidence an issue is the number of shunts.

It also would not surprise me that the connector manufacturer over promised. One thing bugging me here is the engineers at Nvidia are not stupid - they may have been betrayed by a connector that does not live up to its promises. Never forget that there could be several contributors to this problem, I've only been touching on the most obvious ones.

arestavo · Sunday at 7:30 PM

TXE36 said:
... Were 3090 TIs ever made with PCIe 8 pin connectors and not the 12VHPWR? ...

As I understand it, all 3090 TIs had 16 pin power connectors. Same for the 30 series FE cards other than they didn't have the four sense pins, IIRC, making them 12 pin - but the same basic connector other than that.

erek · Sunday at 9:10 PM

i ordered today just now:
WireView Pro II GPU
Alignment: Normal | Variant: Black
Product number: S-TG-WV-P2-H19N

Balkroth · Sunday at 10:10 PM

I've had mine for about 2 weeks, but I haven't installed it yet.

erek · Sunday at 10:48 PM

Balkroth said:
I've had mine for about 2 weeks, but I haven't installed it yet.

why haven't you installed it yet

Balkroth · Sunday at 11:06 PM

erek said:
why haven't you installed it yet

No real reason other than I haven't taken the time to put it in, it's all built and ready and sitting right next to me.

MaZa · 2026-10-19T05:09:37-0400

TXE36 said:
How do you know those connectors, by themselves are good enough? The littlegreenman posted several pictures of burned up PCIe 8 connectors. All the 1080TIs had current balancing and I suspect the practice has gone back a long way. I think all the 3090 TIs had current balancing.

I cannot speak for all the pictures but usually when PCIe 8 connectors burn up the cause is external, something like a crappy power supply blowing up and frying everything or modded cards that got pushed too far. PCIe 8 is tough, tougher than 12whpwr apparently, but nothing is immune to catastrophic part failures nor some score hunter intentionally pushing limits way beyond spec.

TXE36 · 2026-10-19T10:43:14-0400

MaZa said:
I cannot speak for all the pictures but usually when PCIe 8 connectors burn up the cause is external, something like a crappy power supply blowing up and frying everything or modded cards that got pushed too far. PCIe 8 is tough, tougher than 12whpwr apparently, but nothing is immune to catastrophic part failures nor some score hunter intentionally pushing limits way beyond spec.

Agree. I think some if it could be "single rail" power supplies that can drive all their output down a single wire. If that wire somehow gets shorted to ground, the wire will burn up before the power supply overcurrent protection can save it.

A minor quibble, but the terms current balancing or current balancer may be easier to understand if one calls it a load balancer. Both terms really mean the same thing, but load balancer is much more direct, i.e. it balances the load across the incoming power wires.

I don't know if the PCIe 6 pin and 8 pin connectors are robust enough by themselves at high loads without load balancers.

Prior to Pascal, one could BIOS mod more power. Nvidia shut that down by encrypting the BIOS. For Pascal, "shunt modding" was common. Both of these, done without care, can burn up all kinds of stuff, including PCIe 8 connectors. "Score hunting" as you say.

RTX 5090 FE Molten 12VHPWR

Limp Gawd

Limp Gawd

2[H]4U

Attachments

Limp Gawd

Limp Gawd

2[H]4U

[H]F Junkie

Limp Gawd

Limp Gawd

2[H]4U

Limp Gawd

2[H]4U

Limp Gawd

[H]F Junkie

Gawd

Limp Gawd

2[H]4U

2[H]4U

2[H]4U

2[H]4U

2[H]4U

Limp Gawd

2[H]4U

Limp Gawd

Limp Gawd

2[H]4U

2[H]4U

Gawd

Limp Gawd

8-D

Limp Gawd

2[H]4U

Limp Gawd

2[H]4U

8-D

Gawd

8-D

Gawd

2[H]4U

Limp Gawd