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

[taka]

If you have the money for the 5090 then you should get an alerting system based on IR/thermal vision.

 

[Vayra86]

Anyways, Der8auers video now has proven that the FE has a design flaw, because it lacks per pin sensing, which the ASUS Astral 5090 does have and thus can balance the load over all six 12V cables.

Really? Makes you wonder how those 3x8 pin (24 wires) managed all these years...

I'd rather think this proves the cable ain't up to spec

If you have the money for the 5090 then you should get an alerting system based on IR/thermal vision.

Come the 9090 they'll give you a free bunker shelter with your $45,000 enthusiast gpu

Yeah, not a single 4090 build I've put together has melted... That being said after watching the Buildzoid video the only 5090 I'd be comfortable owning is the Astral and I can't believe a company as technically capable as Nvidia doesn't have a failsafe if multiple pins have poor contact.

Yeah feels real premium doesn't it, buying the top end of the stack these days...

 

[N/A]

Looks like there might already be a 2nd occurrence of this issue with a 5090, this time with the stock cable that came with the PSU:

It's a design flaw where the first pin gets more than fair share of amps, half. So we can relax now. It's not a pandemic.

 

[TheDeeGee]

It's a design flaw where the first pin gets more than fair share of amps, half. So we can relax now. It's not a pandemic.

Back to the design table and implent proper circuitry so the GPU knows what each of the six 12V connections is pulling, just like how it was with the 3090 Ti.

 

[Dorek]

Stumbled on this x thing.

 

[TheDeeGee]

Also, remember the GTX 680 with unique stacked 6+6-Pin?

Two of those could have worked for the 4090 and 5090 to save space.

 

[Vayra86]

Only one shunt resistor on the 4090 and 5090.

That's a big mistake. Cables like this WILL have unbalanced loads and the card should be compensating for that. If you fail to balance the current across cables, you get melted cables and fire hazards.

This doesn't look like a connector problem to me. No one can possibly expect a little plastic thing to evenly apply pressure to all 12 pins in all scenarios. The EE should have build some degree of sensing and compensation.

This does make me curious because I really don't know, but how did GPUs do this on pcie cabling then? Its not like we didn't pull 500+W into a GPU before today

I mean sure, I reckon you already have the load divided by three on a 3x8p ? But then there could still be unbalanced load? Also let's appreciate the fact the GPU's not dead. Its just molten at the power socket. Is the problem the uneven draw...because the GPU seems to handle that fine, it just wants the juice. The cause of uneven draw isn't the GPU nor the PSU... they're just sending and receiving the power. In my simple mind that just leaves the cable being insufficient, the size inviting too much variance.

Isn't it also far more economical to just upscale the cable a good bit so you can handle the unbalanced load (and likely get less of it as there is more surface area) instead of desiring even more hardware on a GPU where the desire is to reduce PCB sizes?

 

[ty_ger]

Having more control on load balancing could also improve efficiency and with good quality PSU you can keep those for a decade, paying a few bucks more for better power delivery and security is worth it in my opinion...

In practicality, how would you load-balance the PSU end? Think about it.
You would have to increase or decrease the voltage on various pins to encourage the current to flow or not flow where you need it or where you need less of it. How would you do that? With a resistive load? Terrible. Heat and inefficiency. With a buck/boost converter? More inefficiency, another big stage of parts, and more ripple to filter.

It's easy to just say stuff like that, but when you think of it in practice, it falls aparts. A warning about imbalance is about the best you could do at that stage. It would not be a root solution. The root solution would still need to be the card somehow ensuring that it draws even load. That is a smarter solution because it already has most of the needed circuitry there and is the load which is dynamic.

This does make me curious because I really don't know, but how did GPUs do this on pcie cabling then? Its not like we didn't pull 500+W into a GPU before today

I mean sure, I reckon you already have the load divided by three on a 3x8p ? But then there could still be unbalanced load?

Well, simply, it was a way more robust connector capable of safely delivering way more current than rated. That was usually it. Some cards had fancy features; most not.

Check it out:

TLDW: a single 6-pin connector could deliver enough to 2x 8-pin connectors for the card to run without issue without wires or connections overheating. 8-pin connectors were usually actually rated 150% more than ATX's advertised rating, given the wires were spec'd as needed; and that doesn't even include any additional safety margin engineered into that rating.

 

[bitsandboots]

Also, remember the GTX 680 with unique stacked 6+6-Pin?

Two of those could have worked for the 4090 and 5090 to save space.

wow memory unlocked.
and yeah with as tall as cards are now i imagine there is no real disadvantage to this layout aside from uncommon manufacturing

 

[Dragam1337]

This does make me curious because I really don't know, but how did GPUs do this on pcie cabling then? Its not like we didn't pull 500+W into a GPU before today

I mean sure, I reckon you already have the load divided by three on a 3x8p ? But then there could still be unbalanced load? Also let's appreciate the fact the GPU's not dead. Its just molten at the power socket. Is the problem the uneven draw...because the GPU seems to handle that fine, it just wants the juice. The cause of uneven draw isn't the GPU nor the PSU... they're just sending and receiving the power. In my simple mind that just leaves the cable being insufficient, the size inviting too much variance.

Isn't it also far more economical to just upscale the cable a good bit so you can handle the unbalanced load (and likely get less of it as there is more surface area) instead of desiring even more hardware on a GPU where the desire is to reduce PCB sizes?

You were only pulling 75 watt with 6 pin pcie and 150 watt with 8 pin pcie - if you were pulling 500 watt, it was with 3x 8 pin pcie, like on the asus 3090 strix. 150 watts divided by 12 volts means that the most any of the wires on the pcie 8 pin would get was 12 amps... and that was with a much thicker gauge. So basically a none issue. Debauer was getting 23 amps on a single wire on his 12 pin with 5090 for comparison.

And yes, upscaling the 12 pin connector and cable to double the gauge would have solved it most likely.

 

[FoulOnWhite]

In practicality, how would you load-balance the PSU end? Think about it.
You would have to increase or decrease the voltage on various pins to encourage the current to flow or not flow where you need it or where you need less of it. How would you do that? With a resistive load? Terrible. Heat and inefficiency. With a buck/boost converter? More inefficiency, another big stage of parts, and more ripple to filter.

It's easy to just say stuff like that, but when you think of it in practice, it falls aparts. A warning about imbalance is about the best you could do at that stage. It would not be a root solution. The root solution would still need to be the card somehow ensuring that it draws even load. That is a smarter solution because it already has most of the needed circuitry there and is the load which is dynamic.


Well, simply, it was a way more robust connector capable of safely delivering way more current than rated. That was usually it. Some cards had fancy features; most not.

Check it out:

TLDW: a single 6-pin connector could deliver enough to 2x 8-pin connectors for the card to run without issue without wires or connections overheating. 8-pin connectors were usually actually rated 150% more than ATX's advertised rating, given the wires were spec'd as needed; and that doesn't even include any additional safety margin engineered into that rating.

2 thicker wires instead of 12 thinner.

 

[sbacc]

In practicality, how would you load-balance the PSU end? Think about it.
You would have to increase or decrease the voltage on various pins to encourage the current to flow or not flow where you need it or where you need less of it. How would you do that? With a resistive load? Terrible. Heat and inefficiency. With a buck/boost converter? More inefficiency, another big stage of parts, and more ripple to filter.

It's easy to just say stuff like that, but when you think of it in practice, it falls aparts. A warning about imbalance is about the best you could do at that stage. It would not be a root solution. The root solution would still need to be the card somehow ensuring that it draws even load. That is a smarter solution because it already has most of the needed circuitry there and is the load which is dynamic.

Yeah, I admit I didn't thought that far... Would a system shutting off the power delivery in case of over-current on a pin (or several) on the PSU end be doable ? Or is it just better to force GPU maker (Nvidia) and AIB to do a better job at designing on their end.

 

[Degreco]

Ideally speaking, we need more security margin for current (for 600w we need 8 16AWG wires) and current load management and monitoring circuitry on both end (PSU & GPU). That means we need a new connector and a new standard for PSU as well...

This!

There are so many users on Reddit repeating the third party cables argument. Gosh, I guess only Nvidia can construct proper 12v cables nowadays. Any third party cable, be it from Asus, Corsair or any other vendor is simply not up to the task.

The real problems imho:
1. Blackwell was most probaly not adequately budgeted and rushed in R&D because resources were focused on other products. Don't get fooled by the FE cooler - there is no real innovation in this gen. 9 from 10 USD revenue at Nvidia nowadays comes from the data center business. The gaming department is not that important anymore (CEO likes the stage time in Las Vegas and will still use the spotlight to present a new leather jacket, but the gaming branch is not important anymore).
2. Then it was still based on the same process as the last gen (that was cheaper to order at TSMC, see no.1 and their priorities), so they had to use the old Intel CPU innovation trick to make a difference: push more power to the GPU, getting near the limit of specifications and narrowing the headroom
3. And they do not need to change a thing because of their loyal customer base not aware that things changed (well, I am one of those customers, but starting to believe, I will stay with my 4090 for two more years, skipping the upgrade for the first time in years).
I fear, thats the truth.

 

[TheDeeGee]

wow memory unlocked.
and yeah with as tall as cards are now i imagine there is no real disadvantage to this layout aside from uncommon manufacturing

Actually nothing makes sense with GPU connectors right now.

Like why does a 300+ Watt GPU need 3/4 8-Pins when 12V-2x6 is rated for 600 Watts and has smaller pins.

A stacked 8+8-Pin connector would have worked just fine... And as you said, GPU's are no longer 1-slot anyways, and that stacked connector would be small enough.

The choices that have been made the past years are mind-boggling right now...

 

[sbacc]

This!

There are so many users on Reddit repeating the third party cables argument. Gosh, I guess only Nvidia can construct proper 12v cables nowadays. Any third party cable, be it from Asus, Corsair or any other vendor is simply not up to the task.

The real problems imho:
1. Blackwell was most probaly not adequately budgeted and rushed in R&D because resources were focussed on other products. 9 from 10 USD revenue at Nvidia nowadays comes from the data center business. The gaming department is not that important anymore (CEO likes the stage time in Las Vegas and will still use the spotlight to present a new leather jacket, but the gaming branch is not important anymore).
2. Then it was still based on the same process as the last gen (that was cheaper to order at TSMC, see no.1 and their priorities), so they had to use the old Intel CPU innovation trick to make a difference: push more power to the GPU, getting near the limit of specifications and narrowing the headroom
3. And they do not need to change a thing because of their loyal customer base not aware that things changed (well, I am one of those customers, but starting to believe, I will stay with my 4090 for two more years, skipping the upgrade for the first time in years).
I fear, thats the truth.

Blackwell, honestly is really meh at several level. It really seems Nvidia put minimal effort behind this. All those years people were saying Nvidia will never become complacent like Intel was, but right now with AI and data center being their bread and butter I am not so sure any more.

 

[bitsandboots]

Actually nothing makes sense with GPU connectors right now.

Like why does a 300+ Watt GPU need 3/4 8-Pins when 12V-2x6 is rated for 600 Watts and has smaller pins.

A stacked 8+8-Pin connector would have worked just fine... And as you said, GPU's are no longer 1-slot anyways, and that stacked connector would be small enough.

The choices that have been made the past years are mindboggling right now...

I don't know why, but the answer is nvidia is acting like apple in that they are doing novel packaging solutions for no real reason and at the expense of all else. it's some misplaced desire to make things "look good"
their actions are to the detriment of everyone lately.
put aside the connector for a moment, which appears to be mandated by nvidia upon the board partners
the board partners are having a rough time in other ways. during the 4000 series they complained that nvidia was a bad company to cooperate with because nvidia's FE was competing against them, for less money, and with a superior packaging solution that they did not share.
now on the 5000 series, same story but worse in that the board partners had like 1 month or less in which to make and validate a product before the 5090 release.
i can't think of any reason why nvidia is doing any of this except for that they wished they didn't have board partners anymore, could be the sole producer of cards, and could reduce the bill of materials as much as possible to maximize profit.

its weird because i don't get the feeling AMD is a good company to partner with either, but they don't seem to be trying to harm their partners.

 

[TheDeeGee]

One could also run the 5090 at 100 Watt Power Limit, incase 5 out of 6 cables make a bad connection... lol

What a clusterfuck.

 

[Bomby569]

Really? Makes you wonder how those 3x8 pin (24 wires) managed all these years...

I'd rather think this proves the cable ain't up to spec

the problem is it doesn't make sense, it isn't the cable that decides were the load goes, it's just a cable.

 

[x4it3n]

Here guys, Der8auer has his say:

His conclusion is so ON POINT... Let's hope Nvidia are watching and learning! It is totally unacceptable that more than 2 years after the RTX 4090 release we're still at the same point!
der8auer checked and found out that 1 cable goes up to 23 Amps when others are at 2-3 Amps and they should all be between 6 to 8 Amps max ! What are Nvidia doing ?!!

 

[eidairaman1]

His conclusion is so ON POINT... Let's hope Nvidia are watching and learning! It is totally unacceptable that more than 2 years after the RTX 4090 release we're still at the same point!
der8auer checked and found out that 1 cable goes up to 23 Amps when others are at 2-3 Amps and they should all be between 6 to 8 Amps max ! What are Nvidia doing ?!!

They got away with it by deflecting that even it professionals are plugging them in wrong...

 

[A Computer Guy]

How Nvidia made the 12VHPWR connector even worse.​

This video enhanced my understanding of the problem to the point, that in prior discussions on TPU, why simply doubling the connector won't solve the problem this time around.