5090, nothing learned from the 4090 problems...

[AngleOff66]

An electrical Engineers  take on this. Yes, someone that works in the industry.

https://www.reddit.com/r/pcmasterrace/comments/1io4a67/an_electrical_engineers_take_on_12vhpwr_and/

Technical but still an interesting read.

Edited February 13 by AngleOff66

[LucShep]

9 hours ago, AngleOff66 said:

An electrical Engineers  take on this. Yes, someone that works in the industry.

https://www.reddit.com/r/pcmasterrace/comments/1io4a67/an_electrical_engineers_take_on_12vhpwr_and/

Technical but still an interesting read.

Thanks for sharing!
That's mandatory reading for anyone interested in this issue, for RTX 5090, 5080 and 4090 models, and an excelent take on the subject.
I just hope the guy does not suffer any "retaliation" for speaking so openly.

And before it gets deleted on reddit, quoting his entire post:

Quote

Affectionate-Memory4

An Electrical Engineer's take on 12VHPWR and Nvidia's FE board design

Discussion

 

To get some things out of the way up front, yes, I work for a competitor. I assure you that hasn't affected my opinion in the slightest. I bring this up solely as a chance to educate and perhaps warn users and potential buyers. I used to work in board design for Gigabyte, but this was 17 years ago now, after leaving to pursue my PhD and then the last 13 years have been with Intel foundries and briefly ASML. I have worked on 14nm, 10nm, 4nm, and 2nm processes here at Intel, along with making contributions to Foveros and PowerVia.

Everything here is my own thoughts, opinions, and figures on the situation with 0 input from any part manufacturer or company. This is from one hardware enthusiast to the rest of the enthusiasts. I hate that I have to say all that, but now we all know where we stand.

Background: Other connectors and per-pin ratings.

The 8-pin connector that we all know and love is famously capable of handling significantly more power than it is rated for. With each pin rated to 9A per the spec, each pin can take 108W at 12V, meaning the connector has a huge safety margin. 2.16x to be exact. But that's not all, it can be taken a bit further as discussed here.

The 6-pin is even more overbuilt, with 2 or 3 12V lines of the same connector type, meaning that little 75W connector is able to handle more than its entire rated power on any one of its possibly 3 power pins. You could have 2/3 of a 6-pin doing nothing and it would still have some margin left. In fact, that single-9-amp-line 6-pin would have more margin than 12VHPWR has when fully working, with 1.44x over the 75W.

In fact I am slightly derating them here myself, as many reputable brands now use mini-fit HCS (high-current system), which are good for up to 10A or even a bit more. It may even be possible for an 8-pin to carry its full 12.5A over a single 12V pin with the right connector, but I can't find one rated to a full 13A that is in the exact family used. If anybody knows of one, I do actually want to get some to make a 450W 6-pin. Point is, it's practically impossible for you to get a card with the correct number of 8 and 6-pin connectors to ever melt a connector unless you intentionally mess something up or something goes horrifically wrong.

Connector problems: Over-rated

Now we get in to 12VHPWR. Those smaller pins are not the same mini-fit Jr family from Molex, but the even smaller micro-fit. While 16AWG wires are still able to be used, these connectors are seemingly only found in ratings up to 9A or 8.5A each, so now we get into the problems.

The 8-pin standard asks for 150W at 12V, so 12.5A. Rounding up a bit you might say that it needs 4.5A per pin. With 9-amp connectors, each one is only at half capacity. In a 600W 12VHPWR connector, each pin is being asked for 8.33A already. If you have 8.5A pins, there is functionally no headroom here, and if you have 9A pins, yeah that's not great either. Those 8.5A pins will fail under real-world conditions such as higher ambient temperatures, imperfect surface cleaning, and transient spikes from GPUs. The 9A pins are not much better.

I firmly believe that this is where the problem lies. These pins are at the limit, and the margin of error of as little as 1 sixth of an amp before you max out a pin is far too small for consumer hardware. Safety factor here is abysmal. 9Ax12Vx6pins = 648W, and if using 8.5A pins, 612W. The connector itself is good supposedly for up to 660W, so assuming they are allowing a slight overage on each pin, or have slightly better pins than I can find in 5 minutes on the Molex website, you still only have a safety factor of 1.1x.

Recall that a broken 6-pin with only 1 12V connection could still have up to 1.44x.

It's almost as if this was known about and considered to some extent. Here is a table from the 12VHPWR connector’s sense pin configuration in section 3.3 of Chapter 3 as defined in the PCIe 5.0 add-in card spec of November 2021.

Chart noting the power limits of each configuration of 2 sense pins for the 12VHPWR standard. The open-open case is the minimum, allowing 100W at startup and 150W sustained load. The ground-ground case allows 375W at startup and 600W sustained.

Note that the startup power is much lower than the sustained power after software configuration. What if it didn't go up?

Then, you have 375W max going through this connector, still over 2x an 8-pin, so possibly half the PCB area for cards like a 5090 that would need 4 of them otherwise. 375W at 12V means 31.25A. Let's round that up to 32A, which puts each pin at 5.33A. That's a good amount of headroom. Not as much as the 8-pin, but given the spec now forces higher-quality components than the worst-case 8-pin from the 2000s, and there are probably >9A micro-fit pins out there somewhere, I find this to be acceptable. The 4080 and 5080 and below stay as one-connector cards except for select OC editions which could either have a second 12-pin or gain an 8-pin.

If we use the 648W figure for 6x9-amp pins from above, a 375W rating now has a safety factor of 1.72x. In theory, as few as 4 pins could carry the load, with some headroom left over for a remaining factor of 1.15. This is roughly the same as the safety limit on the worst possible 8-pin with weak little 5-amp pins and 20AWG wires. Even the <profanity>tiest 7A micro-fit connectors I could find would have a safety factor of 1.34x.

The connector itself isn't bad. It is simply rated far too high, leaving little safety factor and thus, little room for error or imperfection. 600W should be treated as the absolute maximum power, with about 375W as a decent rated power limit.

Nvidia's problems (and board parters too): Taking off the guard rails.

Nvidia, as both the only GPU manufacturer currently using this connector and co-sponsor of the standard with Dell, need to take some heat for this, but their board partners are not without some blame either.

Starting with the 3090 FE and 3090ti FE, we can see that clear care was taken to balance the load across the pins of the connector, with 3 pairs selected and current balanced between them. This is classic Nvidia board design for as long as I remember. They used to do very good work on their power delivery in this sense, with my assumption being to set an example for partner boards. They are essentially treating the 12-pin as 3 8-pins in this design, balancing current between them to keep them all within 150W or so.

On both the 3090 and 3090ti FE, each pair of 12V pins has its own shunt resistor to monitor current, and some power switching hardware is present to move what I believe are individual VRM phases between the pairs. I need to probe around on the FE PCB some more that what I can gather from pictures to be sure.

Now we get to the 4090 and 5090 FE boards. Both of them combine all 6 12V pins into a single block, meaning no current balancing can be done between pins or pairs of pins. It is literally impossible for the 4090 and 5090, and I assume lower cards in the lineup using this connector, to balance their load as they lack any means to track beyond full connector current. Part of me wants to question the qualifications of whoever signed off on this, as I've been in their shoes with motherboards. I cannot conceive of a reason to remove a safety feature this evidently critical beyond costs, and those costs are on the order of single-digit dollars per card if not cents at industrial scale. The decision to leave it out for the 50 series after seeing the failures of 4090 cards is particularly egregious, as they now had an undeniable indication that something needed to be changed. Those connectors failed at 3/4 the rated power, and they chose to increase the power going through with no impactful changes to the power circuitry.

ASUS, and perhaps some others I am unaware of, seem to have at least tried to mitigate the danger. ASUS's ROG Astral PCB places a second bank of shunt resistors before the combination of all 12V pins into one big blob, one for each pin. As far as I can tell, they do not have the capacity to actually do anything to move loads between pins, but the card can at least be aware of any danger to both warn the user or perhaps take action itself to prevent damage or danger by power throttling or shutting down. This should be the bare minimum for this connector if any more than the base 375W is to be allowed through the connector.

Active power switching between 2 sets of 3 pins is the next level up, is not terribly hard to do, and would be the minimum I would accept on a card I would personally purchase. 3 by 2 pins appears to be adequate as the 3090FE cards do not appear to fail with such frequency or catastrophic results, and also falls into this category.

Monitoring and switching between all 6 pins should be mandatory for an OC model that intends to exceed 575W at all without a second connector, and personally, I would want that on anything over 500W, so every 5090 and many 4090s. I would still want multiple connectors on a card that goes that high, but that level of protection would at least let me trust a single connector a bit more.

Future actions: Avoid, Return, and Recall

It is my opinion that any card drawing more than the base 375W per 12VHPWR connector should be avoided. Every single-cable 4090 and 5090 is in that mix, and the 5080 is borderline at 360W.

I would like to see any cards without the minimum protections named above recalled as dangerous and potentially faulty. This will not happen without extensive legal action taken against Nvidia and board partners. They see no problem with this until people make it their problem.

If you even suspect your card may be at risk, return it and get your money back. Spend it on something else. You can do a lot with 2 grand and a bit extra. They do not deserve your money if they are going to sell you a potentially dangerous product lacking arguably critical safety mechanisms. Yes that includes AMD and Intel. That goes for any company to be honest.

 

The simple conclusions are (and quoting the author also from one of his following replies in the discussion):  

"Any card drawing more than the base 375W per 12VHPWR connector should be avoided. Every single-cable 4090 and 5090 is in that mix, and the 5080 is borderline at 360W.

Messing up the 4090 is a one-off mistake. It happens. It's not good, but one bad product does not set a pattern. The 5090 is BAD. They knew about the problem from the 4090 already. They had what at least appears to be a working solution from the 3090. They chose not to re-implement that solution after seeing the lack of it cause failures.

So there is no way to safely use a 5090? As far as my opinion goes, no. Unless you cripple it down to 5080 power levels, it is simply too power hungry for this connector. It either needs active load balancing (and it better be good at ~600W) or multiple connectors to brute-force a big safety factor.
The Galax HOF 4090 is actually a good example. 2x 12-pins, so in theory 1320W of power capacity on a 450W card, and if I use the derated connector spec of 375W, that's still 750W. If you find a 5090 like that, only then would I be comfortable running at full TDP."

 

Edited February 13 by LucShep

[LucShep]

igor'sLAB has released an article now too.

Groundhog Day: The 12V2X6, melting contacts and unbalanced loads – what we know and what we don’t know

https://www.igorslab.de/en/groundhog-day-the-12v2x6-melting-contacts-and-unbalanced-loads-what-we-know-and-what-we-dont-know/


 

Edited February 13 by LucShep

[scommander2]

1 hour ago, LucShep said:

gor'sLAB has released an article now too.

Groundhog Day: The 12V2X6, melting contacts and unbalanced loads – what we know and what we don’t know

https://www.igorslab.de/en/groundhog-day-the-12v2x6-melting-contacts-and-unbalanced-loads-what-we-know-and-what-we-dont-know/
 

Thanks and it is very informative...

[BitMaster]

After listening to DerBauer and Buildzoid and now reading Igor, I dare to bet that with those cards you won't achieve lifetimes as long as we used to.

I would be astound if as many (maybe 4080 too)4090 and 5080 and 5090 run after 5 years as 1080ti did.

They seem to be made just to last as long as warranty runs, if you are lucky.

 

It is time to show Nvidia your cold shoulder and maybe cut your own finger a bit for your own sake in the future.

Make your thoughts if you wanna be part of this very one-sided game

 

[LucShep]

1 hour ago, BitMaster said:

After listening to DerBauer and Buildzoid and now reading Igor, I dare to bet that with those cards you won't achieve lifetimes as long as we used to.

I would be astound if as many (maybe 4080 too)4090 and 5080 and 5090 run after 5 years as 1080ti did.

They seem to be made just to last as long as warranty runs, if you are lucky.

 

It is time to show Nvidia your cold shoulder and maybe cut your own finger a bit for your own sake in the future.

Make your thoughts if you wanna be part of this very one-sided game

 

Unfortunately I don't have the money but, even if I had, I wouldn't get a 5090, probably not even a 4090, regardless of "best version" and "bestest PSU and cable" for it.
Some units will last far longer than others, or present less simptoms, but all have this garbage power aspect in the boards, that much we now see (ignore it at your own peril). 

We will see reknowned people in the area showing that it "works just fine", somewhat in contradiction to what Der8auer, IgorsLAB and Buildzoid have shown and/or explained (also the electrical engineer fella in the reddit post).
The problem is, proving that a certain combination of "specific RTX5090 model + specific PSU + specific Cable" is fine, won't acomplish anything of what must be in this subject, which is that every model of these GPUs work fully reliable and safely with all available PSU and cable hardware that is compatible, which as it seems they don't. 

Hey, at least there's one positive side in this  honest working guys like Northwestrepair and KrisFix-Germany will have more work and keep doing what they do best, which is repairing GPUs for people who (perhaps) should have known better. 
 

Edited February 13 by LucShep

[Pilotasso]

note: the founders 5090 doesnt have any power monitoring (as mentioned by der8auer in his overheat videos) but AFAIK the AIB partners do, added to that the later are also providing beefier heatsinks. So I say while the AIB are charging extra (because they get 0$ from the GPU itself) maybe if you really want to have a 5090 and peace of mind maybe its worth to fork out that extra cash...if you dont have an AIB 4090. For what I seen in my limited research (I have a strix 4090) MSI is really scoring major points this time around by having some of the best solutions without being the most expensive by far.

Edited February 13 by Pilotasso

[LucShep]

2 hours ago, Pilotasso said:

note: the founders 5090 doesnt have any power monitoring but the AIB partners do, added to that the later are also providing beefier heatsinks. So I say while the AIB are charging extra (because they get 0$ from the GPU itself) maybe if you really want to have a 5090 and peace of mind maybe its worth to fork out that extra cash...if you dont have an AIB 4090.

We're long past the days of 250W GPUs where pretty much any decent PSU could (and can) easily and safely handle it with the good old trusty double 8-pin connector, from which some users will jump from.

Any of these new 5090s will have spikes reaching (surpassing?) the 600W limit, and beyond the 575W TDP, when at full load.
That's on a single 12VHPWR (or 12v-2x6) connector and cable.

Check the Igor'sLAB review of the 5090FE:
https://www.igorslab.de/en/nvidia-geforce-rtx-5090-founders-edition-review-the-600-watt-powerhouse-in-gaming-and-lab-tests/14/

So, sure, certain AIB models may have the power monitoring issue resolved (as all should have had, IMO).
But they'll still be far too power hungry for their single connector. And still lack the active load balancing.  The concerns and possible issues in the matter will still remain.  

 

Edited February 13 by LucShep

[SharpeXB]

There are plenty of good uses for a burned out 4090. This one is still working perfectly

 

Edited February 13 by SharpeXB

[LucShep]

10 minutes ago, SharpeXB said:

There are plenty of good uses for a burned out 4090. This one is still working perfectly

 

LOL 

As nice as it looks, I think that's worth a try with repairing.  Contact these guys:  Northwestrepair and KrisFix-Germany.

[scommander2]

37 minutes ago, SharpeXB said:

There are plenty of good uses for a burned out 4090. This one is still working perfectly

Yup, it is a good idea to cool it down...

[BitMaster]

Cheers 

LoL, that's a good 1st use case, at  least it's electrically safe.

 

Giggles aside, I cannot understand why Nvidia choose to go this way. They must have known that we will put a pile of sh/t on their name if it happens again...and it does...and here we are,

putting the manure fork to work because they really deserve it.

[okopanja]

Here a quick fix for all 5090 owners

 

[LucShep]

49 minutes ago, okopanja said:

Here a quick fix for all 5090 owners

 

MUHAHAHA 

congrats to the author ... "you win the internet!"  
absolutely friggin brilliant! 
 

Edited February 13 by LucShep

[scommander2]

Still the cable connector issue?

 

[TheBiggerBass]

You should really start to become sceptical if a single 12V unit consumes 600W or more. That's a 50A+(!) max current. This would equal a 5,5kW consumption at 110V or a 11,5kW consumption at 230V. That's more current than your kitchen hotplate needs.

[Nightdare]

Another funny detail

...why do we never see "user error" at the PSU side?

[Nightdare]

1 hour ago, scommander2 said:

Still the cable connector issue?

 

 

Hm, but isn't a 12v that comes with a PSU ALSO a "1st party" cable?

Imagine the pillar -> post

Nividia: "Your connectors melted because you didn't use OUR Cable"

PSU manufacturer "Your connectors melted because you didn't use OUR Cable"

[LucShep]

Was just now watching Paul's Hardware, today's "Probing Paul".  Question/Answer #1 and #2 are Nvidia related (attached vid below is time ready)

When he goes into that small trip down memory lane, and remembering that I just....  eeeeeeeh
 

 

27 minutes ago, Nightdare said:

 

Hm, but isn't a 12v that comes with a PSU ALSO a "1st party" cable?

Imagine the pillar -> post

Nividia: "Your connectors melted because you didn't use OUR Cable"

PSU manufacturer "Your connectors melted because you didn't use OUR Cable"

Every electrical expert:  This stuff is bad like really.  

Tech companies:  It's fine.

Edited February 14 by LucShep

[Panzerlang]

3 hours ago, LucShep said:

Was just now watching Paul's Hardware, today's "Probing Paul".  Question/Answer #1 and #2 are Nvidia related (attached vid below is time ready)

When he goes into that small trip down memory lane, and remembering that I just....  eeeeeeeh
 

 

Every electrical expert:  This stuff is bad like really.  

Tech companies:  It's fine.

Whole video just illustrates how all corporations are anti-humanity. Nothing but lies and obscene greed from the shameless spivs in charge.

[LucShep]

Yep, the memes are deserved.

LOL 

 

7 hours ago, Panzerlang said:

Whole video just illustrates how all corporations are anti-humanity. Nothing but lies and obscene greed from the shameless spivs in charge.

Here's a fun fact - the CEOs (or spivs, depending on POV) of two of the biggest colossi in this industry - and certainly the leading ones in the GPU market - are actually relatives.

https://www.tomshardware.com/news/jensen-huang-and-lisa-su-family-tree-shows-how-closely-they-are-related

 

Edited February 14 by LucShep

[scommander2]

Another video:

It also happened with 5080? 

https://www.tomshardware.com/pc-components/gpus/rtx-5080-power-cable-allegedly-melts-at-psu-redditor-reports-another-50-series-failure

Edited February 14 by scommander2

[LucShep]

5 hours ago, scommander2 said:

Another video:

 

Here we go.
Here's the difference between a propper techtuber and others who are there merely to entertain and profit from the views/clicks.
He's interested in going into the rabbit hole, and prove why this design is hot garbage - which it is.
 

5 hours ago, scommander2 said:

It also happened with 5080? 

https://www.tomshardware.com/pc-components/gpus/rtx-5080-power-cable-allegedly-melts-at-psu-redditor-reports-another-50-series-failure

It could be the case that the RTX5080 is not free of these issues.
It has been suggested that the safe limit is 375W per single conector (12VHPWR or 12V-2x6) and, if true, at ~360W the 5080 is then borderline.
Remember, we have to also keep in mind spikes that may surpass max announced TDP when at full load, and possible manufacturing tolerances.


   

   
 source: reddit.com/r/pcmasterrace/comments/1io4a67/an_electrical_engineers_take_on_12vhpwr_and/

 

Edited February 14 by LucShep

[Aapje]

Yeah, my advice would be to only get a 5090 if you are sure that it is a design that has separated power rails. Definitely not the FE.

If you get a 5080, this is not needed if you undervolt the card. You can typically cut the wattage a lot by doing this. Apparently it undervolts very well and you should easily stay under 300 watts at all times: https://old.reddit.com/r/nvidia/comments/1ikev2d/the_rtx_5080_fe_undervolts_really_well_im/

Although all undervolts are subject to the silicon lottery of course.

[Lange_666]

Then there is also this: