NV 12VHPWR issues revisited

[igor_kavinski]

My vote is just to keep cards under a 350W limit!

Totally agree and AMD seems to be a good citizen in that regard. Well, at least since the RX 7000 series.

 

[igor_kavinski]

Groundhog Day: The 12V2X6, melting contacts and unbalanced loads - what we know and what we don’t know | Page 3 | igor´sLAB

I had actually decided not to write anything more about the 12V2X6 connector (formerly 12VHPWR). But after the numerous reports and discussions on the subject, I decided to take another look at it and…

www.igorslab.de

The values change slightly each time they are plugged in, which indicates the general deficiencies of the plug connection (clamping surface, contact). Added to this is the voltage drop, which also depends on chance.

The shortcomings of the 12VHPWR connector, in particular the uneven current distribution through the cable and connector, can cause unbalanced loads where individual pins are loaded more than others. These local overloads lead to increased contact resistance and heat generation, which under certain conditions can cause thermal damage to contacts and cables. In addition, by dispensing with active balancing and splitting the power supply across several rails in the board topology, NVIDIA has itself abandoned possible protective and corrective measures. As the cards directly take over the faulty distribution of the input side, the power load remains uncontrolled, which can lead to escalation under the wrong conditions.

This situation shows how several factors can interact: The inadequate plug connection as a starting point, the resulting thermal issues as a potential symptom, and the lack of protection measures on the board as an untapped opportunity to remedy the situation. Although such problems do not necessarily have to occur, the system remains susceptible to this concatenation if the load and the external conditions coincide unfavorably.

 

[DrMrLordX]

@igor_kavinski

That would explain the intermittent nature of cable failures. Something as simple as plug/pin wear can change current distribution based on resistance.

 

[DAPUNISHER]

 

[igor_kavinski]

https://videocardz.com/newz/reviewer-records-150c-tempeature-on-12v-2x6-psu-power-cable-connected-to-rtx-5090-graphics-card

It's Inno3D's fault. They should've supplied water-cooled cables with their water-cooled card!

 

[DrMrLordX]

Looks like connector melt on the 5090 could become even worse than on the 4090.

 

[Heartbreaker]

Galax has a 5090 with dual 12 pin connectors in their lab, it's not clear if they will, or are allowed to ship to customers, but they should. It also appears to have 5 shunt resistors for measuring incoming power:

https://videocardz.com/newz/galax-rtx-5090d-hof-xoc-with-two-12v-2x6-connectors-breaks-speed-way-world-record-at-3-27-ghz

 

[MrTeal]

Weird that there's 5, I wonder what the extra one is for? There's only 4 input filtering caps, and 12 incoming 12V lines doesn't divide nicely with 5. Maybe they're grouping into four main 12V rails and just thiefing power from one of them (any measuring through another resistor) for the minor rails.

 

[Heartbreaker]

Slight tangent. I just had this YT story suggested by the algorithm:

It seems that NEMA 14-50 (50 Amps) outlets used to attach home car chargers, are melting.

50 Amps is serious current, where you need perfect connections.

Suggestions are Direct wiring is better, if not, make sure you get an industrial grade outlet.

 

[MrTeal]

Slight tangent. I just had this YT story suggested by the algorithm:

It seems that NEMA 14-50 (50 Amps) outlets used to attach home car chargers, are melting.

50 Amps is serious current, where you need perfect connections.

Suggestions are Direct wiring is better, if not, make sure you get an industrial grade outlet.

Eh... I'd probably direct wire the charger, but 50A isn't that crazy and just seems that way because at 240V it's a lot of power. It's not more than a 5090 draws really. Half the 16 year kids with their new license and and $1000 beater are running straight to Crutchfield to buy 2/0 cable and a couple 2kW amplifiers to pull hundreds of amps and keep their whole neighborhood up for hours at a time and they hook that crap up with nothing but a screwdriver and a dream.

 

[amenx]

Galax has a 5090 with dual 12 pin connectors in their lab, it's not clear if they will, or are allowed to ship to customers, but they should. It also appears to have 5 shunt resistors for measuring incoming power:

https://videocardz.com/newz/galax-rtx-5090d-hof-xoc-with-two-12v-2x6-connectors-breaks-speed-way-world-record-at-3-27-ghz

View attachment 119485

Not a new idea, they've done it before with the 4090. What all AIBs should do with the 5090.

 

[DrMrLordX]

Galax has a 5090 with dual 12 pin connectors in their lab, it's not clear if they will, or are allowed to ship to customers, but they should. It also appears to have 5 shunt resistors for measuring incoming power:

Why would they not be allowed to ship to customers?

 

[Jan Olšan]

All partner cards have to get Nvidia approval. The thing that happened is likely that

1) visuals design guys want 1 connector/cable to look slick and they got the upper bosses on board with that idea (against the engineers, possibly).
2) inevitable safety issues
3) you (Founders edition cards) don't want to look bad/usafe next to partner cards that would put two connectors on their stuff to give better safety margins.
4) disallow partners from making 2xcable cards to "solve" problems 2-3.

 

[coercitiv]

5) creates mayhem among PSU makers, as there most certainly are PSU models out there with a single 12V-2x6 connector that are more than capable of handling the 5090.

Spoiler

The entire marketing pitch for this connector was built around 1x connectivity for anything except extreme builds / overclocking.

 

[Jan Olšan]

Well, those can still throw the spider adapter from 8pins on a double-cable card. Which is what people started recommending for RTX 5090 anyway when the native cables started melting, ironical given how everyone first got new PSUs with native/straight 12+4 cables...

 

[DrMrLordX]

@Jan Olšan @coercitiv

NV's reference design for the 4090Ti already uses two connectors, though. All they had to do was copy that. The only technical reason I've heard as to why they didn't do that is that some designer or another insisted on limited PCB space for a specific airflow/heatsink topology, leading to the single connector at an odd angle.

 

[DrMrLordX]

Ugh

For topics related to melting cables etc skip to the 10:00 mark. Der8auer used AI to translate this video so you sort of have to take him at his word that the AI got everything right, so here is the German video without AI alteration:

You can use Google auto-translate if you want to see the stock footage without any AI manipulation.

In any case, after this NV rep lies his way through the interview, it makes me seriously question the sanity of anyone buying their 50-series cards. Seriously, why?

 

[Ranulf]

Seriously, why?

The big green media machine told them to.

 

[gorobei]

Buildzoid goes over the 12vHighFail design and some history

• load balancing on the psu side is not really a realistic option.
• lumping everything into one rail just makes thing worse as micro differences lead to bigger differentials.
• old vega just put 6 phases on one 8pin and the other 6phases on the 2nd 8pin, the voltage regulation controller sorted out any imbalances (the thing they are designed to do). nv with the 600 series started micro managing the assignment of the vrm to what ever rail was closest to the 'ideal' spec voltage.
• the 8pin has enough safety margin that it doesnt matter if you are pulling too much.

really calls into question what is going on over at nv.

it will really be interesting to see how the fail rate affects the resale value of the 5000 series years later.

 

[Ranulf]

it will really be interesting to see how the fail rate affects the resale value of the 5000 series years later.

I've certainly had little interest in looking at 40 series used but definately will avoid 50 series cards.

 

[aigomorla]

Slight tangent. I just had this YT story suggested by the algorithm:

It seems that NEMA 14-50 (50 Amps) outlets used to attach home car chargers, are melting.

50 Amps is serious current, where you need perfect connections.

Suggestions are Direct wiring is better, if not, make sure you get an industrial grade outlet.

This is why u need to NOT cheap on wire when you do that EV.
Also i don't think many chargers can handle 50Amps except for a few like my Tesla Wall Charger Gen2 that has a 72A max.
Most bottom out at around 40A some even lower to 30A.

But yeah you do not want to strap a 16 gauge wire on a Amperage rating that high.
NEC guidelines require 6 Gauge Copper wire, where as when i had my Wall Charger installed, i made sure my electrician was using #2 Copper wire.

But yeah Nvidia did not learn electrical properly, they used too thin of wire not able to handle current, or they assumed load balancing on the rails, and gloriously failed at coordination.

This reminds me of the old PSU's we had which had either single rails or split rails.
And how Single Rails was preffered when you wanted to overclock a kentsfield as those guys drew so much current off the 12V rail.

 

[igor_kavinski]

So funny that a company with as high a valuation as Nvidia cannot afford a proper electrician for a consulting gig

 

[gorobei]

looks like there were enough questions in the comments that Buildzoid decided to go over how you can divide up the power lines.

he does full card break downs for the vega7, asus 2080, msi 2080, and generally makes fun of nv founders/reference.
then some basic splits math.
a bunch of dumb mosfets and extra controllers to shuffle the power around like a game of whack a mole.

12vhighfail just makes it impossible to safety fuse things in a reasonable way.

 

[DrMrLordX]

12vhighfail just makes it impossible to safety fuse things in a reasonable way.

Could they treat each power pin as its own voltage rail?