It's happening again, melting 12v high pwr connectors

[Dr. Dro]

Ignoring other issues with inline thermistors, all the damage I've seen indicates the issue isn't too little resistance (cables melting from overamperage) but too much -- contact pins mate poorly, and thus become victim to (I^2)r. A 50cm, 12A wire with 20 mOhm resistance can dissipate 3 watts along its entire length , but if that power is all focused on a portion of one pin contact, the surrounding plastic will melt.

Yes, which is why you'd use a PTC thermistor (which increases resistance as temperature increases) instead of the more commonly found NTC thermistor (which reduces resistance as temperature increases), although, I don't know if they can be used to say, cut off the supply entirely. It's probably not possible, and I certainly did not take these other considerations into account - it's just something I've been wondering.

 

[ShrimpBrime]

Always go by the extended load, not the spike.

You can have a spike for 10 seconds within these parameters according to documentation provided earlier in this thread. See pic.

Pin rated 105c max
10 seconds at 260c load for any 1 pin at 12v 9.5a.

Then the plastic melts.

Lol.

Melt down
T-minus 10 seconds
9
8
7
6 5 4 3 2 1

O-zone.

Tin over nickel.

So if the pin heats to 260c, I guesstimate a longer than 10 seconds cool down time with 0 power running across the male/female pins.

The rating of the connector resistance of 10mOhm is the same as that of tin.

Technically, the connector needs a fucking fan to cool the pins with the current configuration. (IMO). Edit, Or perhaps a mini Peltier could keep the pins cool. (Idea only)

 

[EmmanuelMar]

An electrical engineer reviewed several gpus and the asus Astral are the only ones that have at least 3 voltage balancing resistors, the rest of the gpu only one, it is a component that costs cents of a dollar, for this engineer the nvidia connector is a disaster and any gpu that carries that connector with a tdp greater than 350 w claims that it should be withdrawn from the market for being dangerous.
As already demonstrated, this connector can be cut cables and have over amperage in the wires that continue to work at high loads, degrading them in a shorter or longer time depending on the over amperage it carries. When we are talking about 580w consumptions of the 5090 the 600w connector is at the maximum that gives.
Bad connection, a cable or several degraded cables, a psu that has poorly balanced electrical loads would make the disaster fast, the system implemented by hardware and software is not enough and should stop when there is an anomaly of overload in any wire.

A GPU like the 5090 or any other that has peaks of 900w for 9 nanoseconds needs a mandatory 1500w PSU.

 

[Knight47]

I wish, there was a way to make EPS 8-pin connector revision (EPS PCIe v2/2025 or whatever), where someone puts four 12V AWG16 lines in one row, and another four AWG16 lines in bottom (GND), rate whole thing at 200W, add a F+ mark on plastic (for compatibility reasons), and call it day by saving this as "standard requirements".
No pin/key changes, no drama on too low safety margins... it could have been so easy

Note : No key changes, because using "new" connector in old port would cause a dead short to ground on 12V (same thing, if old connector get's used in new ports).
PSU simply won't run, due to simple short (if this kind of user error kills any PC component - PSU's manufacturer should be responsible for paying that RMA LOL). Not counting dummies that force PSU to be on, during such event though.

I assume they did not want to use EPS, because the pinout are different compared to 8-pin a a lot of people would just force it in. Why do we not have reverse polarity protection in PC stuff?
Nvidia could have use the bigger Mega-Fit 12pin that's rated for 23.5A/pin and 12AWG cables https://www.molex.com/en-us/products/part-detail/1700010112https://www.molex.com/en-us/products/part-detail/768250012
or go with Mini-Fit Sr rated for 50A/pin where they could even use 8AWG cable https://www.molex.com/en-us/products/part-detail/439151209 https://www.molex.com/en-us/products/part-detail/439141104

Let's assume this is ASUS - not sure - I do not care for Nvidia

9.46 Amps should be green - when igor posted the correct Specs for the connector. 9.46A is smaller as 9.5 A
When I take this screenshot again - 3 wires are close or over the the max of the 9.5A. I wonder why that ASUS card does not shutdown instantly. Pin 5 is definitely out of range. Pin 6 is up to definition. Pin 4 is questionable.

Comparing again with Igors Spec sheet. Do you really want to use a connector with all 6 wires at the maximum of 9.5 A at 100% of the time while playing games for several hours?

I doubt that ASUS card has any accuracy given for the measurement method. Any true rms multimeter comes with a spec sheet with accuracy. How is that for the ASUS card? You need to factor in the inaccuracy in the measurement for the ASUS card also. Just because you see a current or temperature rating, does not mean it is accurate or correct.

In the Matrix review W1zzard says he cutoff is set to 9.2 A.

 

[Degreco]

In the Matrix review W1zzard says he cutoff is set to 9.2 A.

From what we can see and read now there is no cutoff at all, it's a msgbox warning. Of course nobody knows for sure because nothing is documented.

More importantly, it won’t trigger at 9.2A! If Asus had set the warning threshold that low, Astral owners would be seeing pop-ups constantly. The Astral runs 7%+ higher clocks than the FE and pulls up to 636W (review on this site). Under full load in normal conditions - with properly functioning 12V-2x6 cables, no debris, full cable seating etc. etc. - the amps per pin should naturally fluctuate in max between 8 and 9.5A. A screenshot in this thread shows even 9.7A on the single pin.

If Asus triggered warnings at 9.2A, they’d basically be telling new buyers their Astral is defective right out of the box - at least when stressing the GPU with tests like Furmark 1/2.

So, in my opinion, no 5090 in the current state with the current standard will ever get cut off slightly above max amp per wire - because if it did, you wouldn’t be able to operate the 5090 at all.

 

[R0H1T]

Maybe that was the plan all along? All aboard the 5090Ti hype train

 

[pavle]

nvidia's head is so up someone's a$$ that after at least 2 years they do not admit this connector is only good for 300W (375 max.) and not 600. Walking in own error while investing additional work into it. Pride?

 

[dragontamer5788]

Yes, which is why you'd use a PTC thermistor (which increases resistance as temperature increases) instead of the more commonly found NTC thermistor (which reduces resistance as temperature increases), although, I don't know if they can be used to say, cut off the supply entirely. It's probably not possible, and I certainly did not take these other considerations into account - it's just something I've been wondering.

No one makes PTC thermistors for balancing 8Amp steady state loads across 6x pins. Said thermistors are closer to self resetting fuses. When those things pop their resistance climbs to the point where the circuit is effectively cut off.

Copper naturally rises in resistance the hotter it gets. Parallel copper lines, in theory, are therefore self balancing to some degree.

It makes you wonder what the problem is here exactly.

 

[Dr. Dro]

No one makes PTC thermistors for balancing 8Amp steady state loads across 6x pins. Said thermistors are closer to self resetting fuses. When those things pop their resistance climbs to the point where the circuit is effectively cut off.

Copper naturally rises in resistance the hotter it gets. Parallel copper lines, in theory, are therefore self balancing to some degree.

It makes you wonder what the problem is here exactly.

I see. I won't claim to have any expertise regarding this or knowledge of the availability of such parts, I just wondered why. I suppose the thing is just too bloody overloaded.

 

[eidairaman1]

nvidia's head is so up someone's a$$ that after at least 2 years they do not admit this connector is only good for 300W (375 max.) and not 600. Walking in own error while investing additional work into it. Pride?

Its about money

 

[TimatPSUTest.com]

Adding resistance to the cables does help in current sharing as shown in my simulation posted above. This technique is used in parallel redundant power systems for passive load sharing. It would be interesting for someone to add 40 or 50 cm to the +12 volt wires to test this. However, because the GPU card is a constant power load the sum of the current over the six wires will go up.

When I worked at PC Power & Cooling we had customers that would buy power supplies with a custom harness. We had three Molex TM40 crimp machines with the tooling to crimp a variety of pins including the 16 gage HCS ones. We kept documentation on the crimp machines on what tool to use with each pin and wire size combination. This included number settings for all the adjustments. These settings were determined by working with a Molex rep as well as doing temperature rise testing at maximum load of 13 amps and measuring the millivolt voltage drop. We then used calipers to measure the crimp height and width of the crimp joints. Also, we inspected the crimp dies and replaced them as needed. We also did temperature cycling as well as durability testing. We also saved some to test later to simulate aging. As I recall the currents on an eight pin CPU connector would typically be within 10 percent when measured with a clamp on current meter. To ensure we had authentic Molex parts we always bought from an authorized distributor and did not use a so called “equivalent” part.

To avoid getting stuck with a large quantity of bad parts it's also good idea to avoid crimping wires in large quantities before assembling and testing them.

According to this Amphenol Minitek® 12VHPWR Connector System specification the pin contact resistance is 5m (milliohm) max.

For a 20-inch cable the calculated wire resistance would be about 6.7m and 5m for each pin for a total of 16.7m. This would be about 160 mV voltage drop at 9.5 amps. The ideal way to do this test in production would be to use a data accusation system and log the results to observe trends. Keep in mind durability is rated for 50 cycles. The PCB connector side looks like a solid pin so you can expect it to last longer, maybe 500 cycles.

Below is a partial example of a test fixture. Jump all 12 power wires together and use a 100-watt 0.39-ohm resistor to put just below rated current through them and measure the voltage drop in millivolts. Measure voltage across the resistor and use to calculate current. I would also add one input to measure the total resistance of the sideband wires. This way just 16 inputs would be needed. In theory if there are too many high resistance crimps on the harness the low resistance crimps will fail due to the higher current.

In closing, I think the most likely core cause of 12VHPWR cable failures when most cables work fine is due to manufacturing quality issues. If there is some interest I will design and build a 12VHPWR cable tester.

Timothy Post McGrath

Former Senior Engineer PC Power & Cooling Inc.

 

[_roman_]

If Asus triggered warnings at 9.2A, they’d basically be telling new buyers their Astral is defective right out of the box - at least when stressing the GPU with tests like Furmark 1/2.

Warning: Entertainment

This livestream features the NVIDIA RTX 5090 (specifically, the ASUS RTX 5090 Astral OC model) with an AMD Ryzen 7 9800X3D CPU. We're overclocking the ASUS 5090 Astral on air, but we've shunt modded the GPU to bypass power limits. Today, we're benchmarking in 3D Mark Steel Nomad and Port Royal -- and just for some extra fun, we're pointing a high-end thermal camera at the back of the GPU and the 12VHPWR cables. Maybe it'll melt. Or maybe we'll get a record. This isn't meant to be an overclocking guide, but because we're OCing on air today instead of LN2, it's all take-home information that may help. The biggest challenge will be software.

They made a shunt mod.

Just by remembering from hours ago. The 2mOhm resistor was 8mOhm resistor soldered on top. This should be a resistor value of 1.6mohm according to my laziness and webpage. This is without taking into account the accuracy of those resistors. I do not have the datasheets for those parts. https://www.omnicalculator.com/physics/parallel-resistor

I only watched part of it - I was tired. Maybe someone knows what cables they used.

That cooler is unacceptable what ASUS sells here. From the audible noise on the stream.

 

[Assimilator]

If there is some interest I will design and build a 12VHPWR cable tester.

Brother, this thread is 27 pages and counting, we passed "some interest" about 10 pages back! It would be super cool if you could do this for us, actual science is needed here not just armchair theorising. Perhaps contact GN and see about collaborating with them?

 

[dragontamer5788]

If there is some interest I will design and build a 12VHPWR cable tester.

Not a bad idea! With modern uC parts, its not so hard to build a specialized micro / miliohm meter centered solely around the 5mOhm you talked about.

Constant Current circuit of 100mA over 5mOhm connection is on the order of 500uV measured across the wire+pins. Just 1 digit, or ~3 bits, of precision at that level would likely be sufficient?

There's a whole slew of cheap uCs with 18-bits precision ADC at relatively low speeds (like 100Hz amplifiers but you know... tradeoffs and all). There's also ~12-bit ADCs with 16x gain amplifiers on singular $1 uCs (The good ol 8-bit AVR EA or AVR EB chips) with noise-characteristics that should be sufficient for ~500uV region with enough sampling / low-enough frequency. 100Hz is probably fast enough for all of this (I severely doubt that a few dozen cm wire is handling kHz or MHz signals).

-------------

In all honesty, I'd just have 6x parallel AVR EA microcontrollers + shunt resistors measuring all 6 wires concurrently. Plus a "master" microcontroller that can send a specified current that auto-calibrates the 6x individual uCs. Such specialized equipment will likely be what? $100? $200? Probably more? You can afford $6 in uCs (80-cents per AVR EB: https://www.digikey.com/en/products/detail/microchip-technology/AVR16EB14T-I-SL/23029576)

EDIT: Drat. uV of noise is 500uV for the PGA. I admit I'm doing "hard mode" with this design but I kind of wanted to see if testing the individual wires was possible. Its still awesome that pure-digital logic can get this close though in the year 2025.

 

[Endymio]

I think the most likely core cause of 12VHPWR cable failures when most cables work fine is due to manufacturing quality issues. If there is some interest I will design and build a 12VHPWR cable tester.

While this isn't entirely unhelpful, if the connector pins are poor quality or crimped badly, then testing will be extremely difficult. Plug it in once and it may be perfect; unmate and replug, and it may show far below spec.

A GPU like the 5090 or any other that has peaks of 900w for 9 nanoseconds needs a mandatory 1500w PSU.

Nanosecond-width pulses won't cause a PSU to overheat or fail. It might cause supplied voltage to drop and thus be a stability issue -- but in reality any decent PSU's capacitors alone should be able to easily supply double the rated power for pulses of this duration.

 

[Kapone33]

How many years has this issue has been brought up on Social Media?

 

[Dredi]

According to my testing of two different new and like new cables probably manufactured 1 to 1,5 year apart from each other, the plugs on the earlier looking worse (soft plastic bodies, contact tubes moving in the plug body a bit) the current measured with 400W load was on all 6 wires from 5,6 to 5,8 A (at 575W it would be 8,05 - 8,34 A) - I found no problem.

What is your belief that the problem exists based on? Probably not your own measurement. Is it based on DerBauers measurement of his faulty cable - that should not be used at all? That was not a manufacturing error, he destroyed it by misusing it.

Every electrical appliance should be used ONLY with a cable in good condition. That is a common rule and every common man is expected to follow it in his daily life. Any damage caused by a faulty cable would be blamed on the user or installer. Right?

So how can anyone get an idea he can make any judgements based on the premise, that an installer makes a grave error of using a faulty cable?

Every cable supplied with a PSU of reputable brand should be tested in manufacture and should be fault free when new.

What if the problems caused by new cables supplied with PSUs are so rare, that all other problems that may occur in the GPUs (as electronic component failures, etc) are more prevalent? What would be the meaning of all this circus and scandal then?

The problem is that the spec permits cables which you would deem unsafe.

 

[dirtyferret]

posted over the weekend

 

[Degreco]

They made a shunt mod.

LOL, great way not to contribute to the issues at hand. Well, GN is certainly aware of the thing - and Steve said in the beginning of the stream they are not testing it with this stunt (23:53). They just want that 3DMark record. I can already see some people thinking everything is good and proven to be safe by this exercise - just by looking at the comments at YouTube.

But after this last week, a contribution like this feels like a deviation. I mean having an Astral at hands, the only card with a readout on the pins and then doing a shunt mod on this card to break 3DMark records? Vacation week at GN. It's a fun distraction. I’d even call it a ‘commercial’ for Asus/Nvidia in this context - but it’s still Steve, and I respect him.

Joe Stepongzi then gives it away... Steve says something about either the cable burns or they get a record today. Stepongzi right after that part with the cable burn mumbles "good possibility"

____________
Update: Just downloaded the transcription (no way I will watch a nearly 4 hour stream) and asked ChatGPT to find hints on the connector issue. It found a little passage telling a lot:
2:19:18 - "and thanks to a combination of the wire view and the Asus rear fan spraying air all over the cable [and] the connector I think uh you know our title is overclocking until we melt"[..]

2:19:32- [..]"cause there is too much cooling that connector. Also it's just not a long enough test" - "yeah I don't think it's going to be saturating it long enough to make a difference" - "needs to be like a Furmark load yeah"

Yes, GN knows

 

[BoggledBeagle]

The problem is that the spec permits cables which you would deem unsafe.

What???

If there are some unsafe cables, they are not safe because the connectors in the plugs due to damage or manufacturing flaws do not have enough grip on the pins.

The connector tubes in the plugs may be too loose or they are shifted back in the plug body.

 

[Degreco]

More fodder. 4d old now, I somehow missed it. Just for information.

Pin loads with with 2 different Cablemod cables on the Asus Astral 5090:

Cable 1:

Cable 2:

He says, cables were used a lot (approx 30 seatings).

• Brand-new native PSU cable + brand-new Nvidia adapter: All pins green and load balance ok
• He then bought a new Cablemod cable: All pins green and load balance ok

Even slightly bending cables changes the result (assume slightly worsens it towards unbalanced loads).
Slightly fiddling with the connector => Results change again.
Whole reseating => Results change again.

His verdict at min 7:48: "So I am gonna say in any single one of my reviews because i have seen this: If you're spending that much money on a 5090 you owe it to yourself to get a new cable"

"If you've got a new PSU going in, you are not going to have a problem" (use the new native one provided with the PSU)


Interestingly, I looked into the ASUS GPU Tweak III FAQ - I wanted to know what they recommend if a Pin is red (load balance problem):
https://www.asus.com/support/faq/1048435/

So, Asus recommends not to change the cable, but instead to reseat it (which feels not like a good solution).
Still, I think at this point you should probably use new cables just to be sure. And then try not to fiddle with it too much. This toy connector with its miniature wires and pins operates at the limit.

On a sidenote: Let's talk about the definition of 'immediate notification' with the Astral 5090 - this is a very hidden notification as of now

 

[x4it3n]

De8auer is using 2x8 pin PCIe which are meant to be rated @ 150W rather than the 300W that he is trying to pull through them.
The bundled adaptor has 4x8 pin PCIe. Not that it should be pulling 23 Amps through one wire tho.

Sure but the 8-pin connectors can handle 300W even though they're rated at 150W, whereas the 12VHPWR/12V-2x6 connectors are rated for 600W and the max they can handle is supposed to be between 600W and 660W which is ridiculous safety-wise!
Those cables should have been rated for 400W max and NVIDIA should have used 2x 16-pin connectors on their 5090s + AIBs ! The fact that even AIBs cannot have 2x 16-pin connectors for their Custom OC models is beyond me and doesn't make any sense (except making people want to buy the FE since there is not much to get out from AIB OC models anymore...).

Maybe that was the plan all along? All aboard the 5090Ti hype train

If there is ever a 5090 Ti (which I highly doubt) it will for sure have 2x 16-pin connectors because there's no way they can go around that...

 

[Bomby569]

it's insane to me that Nvidia continues to move forward on this and gets no actual pushback. All the influencers that destroyed Intel with such ease must really like Nvidia wink wink

 

[Endymio]

it's insane to me that Nvidia continues to move forward on this and gets no actual pushback. All the influencers that destroyed Intel with such ease must really like Nvidia wink wink

Perhaps because most 5090 purchasers are able to think for themselves and look at the actual data, rather than buying into "influencer" generated hype.

 

[Bomby569]

Perhaps because most 5090 purchasers are able to think for themselves and look at the actual data, rather than buying into "influencer" generated hype.

most people that buy 5090's (or 5080's or 5070ti's) for those prices would buy whatever crap Nvidia threw at them, they are just the Apple buyers of the PC world right now.

Nvidia please i can't wait for the next overpriced crap you make, i will buy one of whatever every generation