News Technician Repairs Hundreds of RTX 4090s With Melted Power Connectors Every Month

[3ogdy]

I may not be good with math, but 20 - 25 a week simply doesn't amount to "hundreds every month". It amounts to a hundred at best.

 

[George³]

I may not be good with math, but 20 - 25 a week simply doesn't amount to "hundreds every month". It amounts to a hundred at best.

56.53 weeks. From release date of RTX 4090. 56.53*25=1413 failed cards in this service as theoretical maximum.

 

[Roland Of Gilead]

I'm wondering why PCI Sig isn't in this equation more. Not vendor, not user, but designer flawed. It's passably it could be a bad implementation by nVidia, but given PCI Sig have changed the design, says to me at least, that they can be apportioned blame too.

 

[vanadiel007]

Nvidia is not a power supply vendor. Can we please hold down the conspiracy theories?

I never said they were. Nvidia is simply part of the system where manufacturers make changes every so often, to increase component sales.

I do like the idea of making the power connector part of the card and motherboard, so no more cables to the GPU are needed. I think that is a real improvement.

But replacing 3 individual connectors with a single connector, that's just silly.

 

[B Heuer]

Doing simple math and using a wire gauge chart for low voltage applications shows they connector is deeply flawed. It has no margin baked in for overdraw which a loose connection can cause. Remember when doing that calculation you have to convert smaller wires to their single wire equivalent.

So 450watts at 12v is going to pull 37.5a. Copper at over 60c operational temp needs to be 6 gauge to handle that kind of load. As 8 gauge only leaves you 3 amps of wiggle room and only if your wire temp is over 60c. Wire gauge goes up as operational temps go down as well so you spec over gauge for that. It takes (16) 18 gauge wires to be the equivalent of 6 gauge. Which is the standard size wire for PCIe connectors.

Nvidia in their infinite wisdom thought they could break simple physics. The new 12 pin connector is only using (12) 18 gauge wires instead of the 16 needed to hit the right gauge of wire for their 450w max. Their current spec puts them at 7 gauge not including the smaller 4 sense pins.

They didn't leave any room for their max load while people are seeing some after market cards pulling 600 watts. The connector is woefully under designed. Older cards often had a third 8 pin despite it not being necessary. Which would bring the effective wire size to 4 gauge and they drew a lot less power than 4090s. 3080ti at 350watts draws 29 amps and only requires 8 gauge to have a good margin yet they gave you an effective gauge of 4 with (3) 8 pin connectors.

You can see the stark comaprison to the 4090 which is 1 gauge under its minimum safe wire size and asking 4 tiny wires to pick up the slack. It's no wonder a loose connection is resulting in the temps it is and causing melting of the connector to occur. The wires simply can't handle the amp load much less the extra impedence from a bad connection.

 

[bill001g]

Doing simple math and using a wire gauge chart for low voltage applications shows they connector is deeply flawed. It has no margin baked in for overdraw which a loose connection can cause. Remember when doing that calculation you have to convert smaller wires to their single wire equivalent.

So 450watts at 12v is going to pull 37.5a. Copper at over 60c operational temp needs to be 6 gauge to handle that kind of load. As 8 gauge only leaves you 3 amps of wiggle room and only if your wire temp is over 60c. Wire gauge goes up as operational temps go down as well so you spec over gauge for that. It takes (16) 18 gauge wires to be the equivalent of 6 gauge. Which is the standard size wire for PCIe connectors.

Nvidia in their infinite wisdom thought they could break simple physics. The new 12 pin connector is only using (12) 18 gauge wires instead of the 16 needed to hit the right gauge of wire for their 450w max. Their current spec puts them at 7 gauge not including the smaller 4 sense pins.

They didn't leave any room for their max load while people are seeing some after market cards pulling 600 watts. The connector is woefully under designed. Older cards often had a third 8 pin despite it not being necessary. Which would bring the effective wire size to 4 gauge and they drew a lot less power than 4090s. 3080ti at 350watts draws 29 amps and only requires 8 gauge to have a good margin yet they gave you an effective gauge of 4 with (3) 8 pin connectors.

You can see the stark comaprison to the 4090 which is 1 gauge under its minimum safe wire size and asking 4 tiny wires to pick up the slack. It's no wonder a loose connection is resulting in the temps it is and causing melting of the connector to occur. The wires simply can't handle the amp load much less the extra impedence from a bad connection.

It is not the wire that is the problem it physical connector. If you read the specs from pci-sig where this connector is defines it requires that each pin be able to carry 9.2 amps which gives you 662 watts at 12 volts. So there is 10% overhead built in.

You can assume that all the power supply makers combined know all about wire gauge and know 18 gauge wire can carry the proper current over just 6 12 volt leads.

If the problem was even close to as bad as you imply every single card would immediately burn up. The vast majority of the cards work fine. The wires do not even get warm.

It is all related to the plug for whatever reason sometime not making proper contact. This leads to all that power now trying to flow over fewer pins and it exceeds the 9.2 amps rating of one or more of the pins.

 

[tamalero]

the real problem is a GPU that uses too much power

This!
I have one of those limited to 450W and well. I have not seen any issue at all so far.. and I live in a very hot environment.

 

[tamalero]

It is not the wire that is the problem it physical connector. If you read the specs from pci-sig where this connector is defines it requires that each pin be able to carry 9.2 amps which gives you 662 watts at 12 volts. So there is 10% overhead built in.

You can assume that all the power supply makers combined know all about wire gauge and know 18 gauge wire can carry the proper current over just 6 12 volt leads.

If the problem was even close to as bad as you imply every single card would immediately burn up. The vast majority of the cards work fine. The wires do not even get warm.

It is all related to the plug for whatever reason sometime not making proper contact. This leads to all that power now trying to flow over fewer pins and it exceeds the 9.2 amps rating of one or more of the pins.

Didn't Steve also tested that sometimes the connector pulls more power from certain cables and not well balanced between all cables?

 

[Wuwu]

 

[LabRat 891]

Didn't Steve also tested that sometimes the connector pulls more power from certain cables and not well balanced between all cables?

That's the issue w/ parallel conductor LVHC:
If the resistance on one single connection increases, you can start a 'thermal runaway'.

If an individual connection/conductor heats up, its resistance increases, putting more 'load' onto the other conductors.
The others begin to heat and increase in resistance; before you know it there's melting and/or fire.

Much less of a problem in PCs (vs. Automotive and Aerospace) but, that is why you don't (often) see 'paralleled conductors' for things like Starter motors, electric heaters, etc.



Something I'd like to see tested (that's yet been mentioned anywhere, AFAIK) is Inductive Kickback (also thermal-expansion cycling)
from rapid changes in current.

I wonder if these plugs can 'jiggle' themselves loose under rapidly-varying high current loads?
Ex. a 900W transient load would be 75A, and (as mentioned) may not be evenly distributed amongst all conductors.

 

[3ogdy]

56.53 weeks. From release date of RTX 4090. 56.53*25=1413 failed cards in this service as theoretical maximum.

Sure, but it's still not hundreds of GPUs a month.

 

[George³]

Sure, but it's still not hundreds of GPUs a month.

Perhaps the title is a little too much. When it specifically applies to this electronics repair shop only. If we zoom in to the size of the planet, maybe the monthly number is 10000? The number is a completely arbitrary guess as there is no public data from any other service.

 

[Wuwu]

The RTX 4090 came out 13 months ago. AFAIK all 4090s have at least 12 months of warranty; many have more, e.g. those bought direct from Nvidia have 3 years.

So is this guy getting all these damaged cards just from people who bought 4090s in the first month it was out, and models with the shortest warranty? Or is he dealing specifically with people who have done something to void their warranty? Something about this doesn't quite add up.

Edit: On 2nd thought, he implies he has been fixing these for some time, so he must have been dealing with cards that are within their warranty period. Why would these people be sending their cards to independent repair shops while they're still under warranty?

As for the question as why people do X vs Y, why do you think IQ tests exist?

 

[Wuwu]

I wouldn't pay this dude too much mind. He's just out to drive clicks to his site and promote himself. He's been posting 4090 videos for a while and been called out by Gamers Nexus for misleading data in his videos. Most of his videos just look like informercials for himself. Everyone else is wrong, everyone else does terrible work. Only I know the truth and how to fix things. One of them he actually holds up a 4080 PCB and calls it a 4090. He wants to call out everyone else but can't identify a 4080 vs a 4090 PCB that he's holding?

Since when has Gamers Nexus repaired RTX 4000 series GPU's?

I too would be a bit dramatic if I spent £1700 on a GPU for it to melt it's self down.

I can maybe understand doing that with metals to make something, but GPU's tend to have a singular purpose.. ya know?

 

[Wuwu]

I never said they were. Nvidia is simply part of the system where manufacturers make changes every so often, to increase component sales.

I do like the idea of making the power connector part of the card and motherboard, so no more cables to the GPU are needed. I think that is a real improvement.

But replacing 3 individual connectors with a single connector, that's just silly.

Here you are expressing the theory vs the reality, respecting the nature of which reality works.. example it is stupid to do it just to say, hey look what cool think I did as a manufacturer.

 

[Moonstick2]

All the comments saying it's wholly/partially user error keeps making me think of Apple and "you're holding it wrong".

 

[MJS WARLORD]

I have seen videos of wires and plugs melting on these cards and as i was planning on getting a new rig built i contacted the company who was going to build the rig for me. They said they were aware of the problem and they throw away the cables that come with the card and out source replacements as they are better quality.

They took the decision to do this because if the cards cause a rig to fail they would be legally obliged to repair or replace everything they got damaged.

 

[Order 66]

I have seen videos of wires and plugs melting on these cards and as i was planning on getting a new rig built i contacted the company who was going to build the rig for me. They said they were aware of the problem and they throw away the cables that come with the card and out source replacements as they are better quality.

They took the decision to do this because if the cards cause a rig to fail they would be legally obliged to repair or replace everything they got damaged.

which is exactly what Nvidia should do as well.

 

[Bamda]

Hey, living on the edge has its benefits.

 

[Order 66]

Hey, living on the edge has its benefits.

Like what, in this context? the only benefit I can see would be a new RTX 4090 whenever it melts, but that is a hassle having to RMA it.

 

[tamalero]

Sure, but it's still not hundreds of GPUs a month.

For one dude is still a lot.
Who knows how many are being returned to the manufacturers.

 

[tamalero]

I have seen videos of wires and plugs melting on these cards and as i was planning on getting a new rig built i contacted the company who was going to build the rig for me. They said they were aware of the problem and they throw away the cables that come with the card and out source replacements as they are better quality.

They took the decision to do this because if the cards cause a rig to fail they would be legally obliged to repair or replace everything they got damaged.

Anyone remembers the fire & melting video card PCIE VGA riser extension cable?
And how much it took for them to rectify?

 

[tamalero]

Like what, in this context? the only benefit I can see would be a new RTX 4090 whenever it melts, but that is a hassle having to RMA it.

I wonder if they would send you only refurbished instead of new XD

 

[vanadiel007]

Here you are expressing the theory vs the reality, respecting the nature of which reality works.. example it is stupid to do it just to say, hey look what cool think I did as a manufacturer.

Think about it:

1. Who all has the existing connectors? Almost every computer sold in the past 10 years.
2. Who has the new connector? Nobody (yet).
3. What happens when we introduce a new connector? Sales of power supplies, profit.

Remember, the new connector does not really offer any practical improvements over the old connector. In fact, one could argue it's harder to manage the cabling because it's so bulky.

 

[TJ Hooker]

That's not the argument. The initial design of the new connector caused notably higher rates of failure than the old 6/8 pin. Further, a few relatively simple changes to the new design resolved some very obvious shortcomings in that initial design. The changes are not limited to helping just with improperly seated connectors. That said, your assertion that improperly seated connectors are automatically user error is false. The initial design is bad enough that improperly seating can more easily happen.

The only changes made for 12V-2x6 are to shorten/retract the sense pins, such that they are last contact/first break. How does that help with anything other than improperly seated connectors?