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

[taka]

Instead of wasting R&D dollars on cables and connectors, Nvidia should be focusing on reducing power consumption in GPUs without losing or lowering their graphical abilities.
Ideally they should be no higher than 300W.

Nvidia already killed "Moore's law". Jensen said that on 40 series launch. Now Nvidia is at war with "Ohm's law". Next is Kirkoff, Newton, Joules and Pythagoras.

 

[BoggledBeagle]

It's a called a connector for reason, you can disconnect and connect it again. This must last more than 1 cycle lol.

Yes, but you must use metal that holds shape for that, and not mud metal. Mud metal is on connector manufacturers.

This 12pin piece of shit connector is so living on the edge that it hurts.

They are forcing it to live like that. On a 5090 it should have a buddy on board.

What some of you here seem to not comprehend is a fact, that there should be no place for error (any error, not just user error) when talking about this poorly dimensioned connector. It has so thin pins and low mating surface that there should be proper locking mechanism as part of connector itself.

Yes, see the plug development thread here:

Proposed new Power Connector

With all the attention on melting graphics card connectors, this might be an opportunity to switch manufacturers to a connector that we like. Here is my proposal - please comment whether you like the idea or not, any changes and if we get enough people behind it we'll try to make them listen...

www.techpowerup.com

Tabs are much better than tiny pins.

 

[Bomby569]

But the only reason that on wire is taking all the load is because it's resistance is significantly less than the others; that's either on the GPU end (testable from the connector to the shunt resistor, if it's consistent across all 6 12v pins then it's NOT the board.

And if it's not the board, you are limited to the fault being with either the cable or the connection.

i have some problem blaming this on the cables, after all it's just wires, what do you want to do here? and all types of cables failed but only failed on 4090 and 5090, so maybe there is a clue here

 

[deathjester]

i have some problem blaming this on the cables, after all it's just wires, what do you want to do here? and all types of cables failed but only failed on 4090 and 5090, so maybe there is a clue here

Why?

This is how the 5090 FE wiring works:

So the PSU has 6 connectors passing 12V to the 5090, from a single 12V power rail.

The 5090 then combines all of that back together into rail before it hits the single shunt resistor on the card.

The difference in resistance between 23 amps (0.5 Ohms) and the lowest - 2 amps, (6 Ohms) is equivalent to over 1000 feet of AWG 16, so it would be hard (without a serious manufacturing issue on the PCB it's self) to have this huge increase in resistance. This should also be testable as you can test the resistance from the 12V-2x6 connector on the GPU to the single shunt resistor with a multimeter.

If you aren't seeing this significant increase in resistance between each of the pins, then the fault CANNOT be with the GPU specifically, and must be related to some combination of connection quality across the connectors.

 

[JAS_design]

Yes, but you must use metal that holds shape for that, and not mud metal. Mud metal is on connector manufacturers.

This is a valid point. I mentioned earlier that I use Molex MicroFit 3.0 (which is roughly equivalent to the connectors used on GPUs) for my own projects. I only ever use official Molex crimp contacts, and those are made from phosphor bronze, which is much more resistent to deformation than the plain brass (or worse) used by some third-party contact manufacturers.

Even the best crimp contacts of this type are only rated for about 30 cycles before performance can no longer be guaranteed. With third-party contacts, who knows what you're really getting.

 

[Dredi]

Andres Schilling did a bunch of tests with many different psus/build types and 5090 versions including AIBS and had no issues.

If you check the initial values he got and plot the numbers to a current divider calculator, you get unsafe resistance distribution. He just had bad connection on one pin instead of maybe three wich would be enough to get to derbauer numbers.

The difference in resistance between 23 amps (0.5 Ohms) and the lowest - 2 amps, (6 Ohms) is equivalent to over 1000 feet of AWG 16, so it would be hard (without a serious manufacturing issue on the PCB it's self) to have this huge increase in resistance.

It does not need to be a huge increase though.

If you calculate with 0.5 milli Ohms and 6 milli Ohms you get the same distribution of current!

 

[Zach_01]

Any one cares to wonder what those differences will become overtime?
Like after a few months or a year? Or just unplug it and re-plug it only a couple of times because you want to clean something?

 

[Dredi]

Any one cares to wonder what those differences will become overtime?
Like after a few months or a year? Or just unplug it and re-plug it only a couple of times because you want to clean something?

View attachment 384474

The last of those has a resistance 26 times higher on one lead compared to another. Things would be pretty bad already if two more would get similarly bad contact.

 

[jnv11]

I noticed that there is an argument on whether the problem is due to a cable fault, user error, a plug fault, or a socket fault. Have you thought of a design error behind the socket? Buildzoid's video below makes a case that a circuit design error (either weak or no current balancing across the pins in the 12V-2x6 or 12VHPWR connection) on the video card behind the 12VHPWR socket or 12V-2x6 socket is likely to be the main problem instead of the problems being argued about in earlier posts.

 

[BoggledBeagle]

This is a valid point. I mentioned earlier that I use Molex MicroFit 3.0 (which is roughly equivalent to the connectors used on GPUs) for my own projects. I only ever use official Molex crimp contacts, and those are made from phosphor bronze, which is much more resistent to deformation than the plain brass (or worse) used by some third-party contact manufacturers.

Even the best crimp contacts of this type are only rated for about 30 cycles before performance can no longer be guaranteed. With third-party contacts, who knows what you're really getting.

I am afraid that sourcing a properly made connector from the proper metal and ensuring that you are still getting what you ordered may be quite difficult for the manufacturers even today, and almost impossible few years ago when these connectors started.

I believe that there are some cables out there which are really almost 1 use only, or which should not be used at all.

DerBauer video is illustrating how negligent he is in misusing a cable, which has been supplied as an accessory for a powersupply for 1 pc build, not controlling its shape while he was misusing it and almost causing fire by using it. I am not sure why he published it. He made a fool out of himself.

 

[tugrul_SIMD]

Easy solution: liquid metal for conduction in connectors. Dont forget to use duct tape.

 

[qxp]

Given how the electronics are configured (very limited with only one circuit and combining the load on the GPU side, with only a single shunt resistor) there has to be some difference in the resistance down each of the connections.
Since any potential distance is pretty much irrelevant (as it's going to be down to less than 2%) then there must be an alternative source that is increasing resistance in the setup.
23 Amps - 0.5 ohms (282W)
11 Amps - 1.09 ohms (132W)
8 Amps - 1.5 ohms (96W)
5 Amps - 2.5 ohms (60W)
3 Amps - 4 ohms (36W)
2 Amps - 6 ohms (24W)

Which works out to 630W total, so some of those figures are probably slightly high, but this will be down to fluctuations in the readings.

You're looking at 12x the resistance between the lowest usage and the highest usage, so if there is no issue with the cable, and no issue with the 12V-2x6 connection then it should be fairly easy to test the resistance from the 12V-2x6 connector to the shunt resistor for each of the 12V pins.
Bare in mind here that 4 ohms of resistance is equivalent to around 1000ft of AWG 16 wire, which while not massive is really significant in the above example.

If you can see this difference on the PCB between the shunt resistor and the 12V-2x6 connector then it's a manufacturing issue.

If its not there then it's either the connector which is incorrectly seated creating resistance, or an issue with the cable (unlikely). Especially when other builders are not seeing the same issue.

This is a good argument, but you need to scale down your resistances by a factor of 10. A contact can have resistance of mOhm's to 0.1 Ohm, more if there is heavy oxidation, or surface contamination. Surface contamination could be usual dirt, but could also be outgassing from plastic. If you have a piece of rubber nearby there might be sulphur on the surface. You don't need a lot to change contact resistance as it is a surface effect after all.

 

[deathjester]

This is a good argument, but you need to scale down your resistances by a factor of 10. A contact can have resistance of mOhm's to 0.1 Ohm, more if there is heavy oxidation, or surface contamination. Surface contamination could be usual dirt, but could also be outgassing from plastic. If you have a piece of rubber nearby there might be sulphur on the surface. You don't need a lot to change contact resistance as it is a surface effect after all.

Do I?

 

[Dredi]

This is a good argument, but you need to scale down your resistances by a factor of 10. A contact can have resistance of mOhm's to 0.1 Ohm, more if there is heavy oxidation, or surface contamination. Surface contamination could be usual dirt, but could also be outgassing from plastic. If you have a piece of rubber nearby there might be sulphur on the surface. You don't need a lot to change contact resistance as it is a surface effect after all.

The current distribution would be the same, even if ’m’ was written before the Ohm in those calculations.