Follow Slashdot stories on Twitter

 



Forgot your password?
typodupeerror
×
IT

Nvidia RTX 4090's Power Draw May Be Too Much For Its Power Connector To Handle (arstechnica.com) 118

An anonymous reader shares a report: Nvidia's $1,599 GeForce RTX 4090 is an incredibly powerful graphics card, but its performance comes at the cost of high power draw. Like a few of the RTX 3000-series cards, Nvidia uses a new kind of 16-pin 12VHPWR power connector to supply all that power to the card -- you can plug up to four 8-pin GPU power cables into the 12VHPWR adapter, which then plugs into the connector on the GPU, saving some board space.

But at least two RT 4090 users are now reporting that their 12VHPWR connectors have overheated and melted during use. These complaints are sourced from Reddit (via Tom's Hardware), so take them with a grain of salt -- we don't know the exact configuration of either user's PC setup. The specific model of graphics card (a Gigabyte RTX 4090 Gaming OC for one user, an Asus RTX 4090 TUF Gaming OC Edition for the other), the power supply, and any number of other factors could have contributed to the connectors overheating.

This discussion has been archived. No new comments can be posted.

Nvidia RTX 4090's Power Draw May Be Too Much For Its Power Connector To Handle

Comments Filter:
  • Deja Vu. (Score:4, Insightful)

    by Ostracus ( 1354233 ) on Tuesday October 25, 2022 @08:30PM (#62998235) Journal

    Gamers Nexus covered this. [youtu.be]

    • So did JayzTwoCents. [youtu.be] Good information on how it might be the bend radius of the new 600Watt cables that are causing insufficient pin contact for the amount of current each pin is supplying.

      • Re:Deja Vu. (Score:4, Insightful)

        by thegarbz ( 1787294 ) on Wednesday October 26, 2022 @01:40AM (#62998703)

        Why on earth we insist and putting the connector in a position where such a bend is required is beyond me. GPUs are big enough that in many cases there's no option other than breaching this bend radius. It would make more sense to have the connector come with right/angle connectors or be in a different location on the card (such as the back, which used to be a problem when our PC cases were full of HDD and GPUs were insanely long, but they've gotten shorter and fatter).

        • Because on the back it would run up against another big thing in computer cases. The CPU cooler which can be pretty big. Memory and M.2 SSDs are also in the general area with their own cooling issues.

          • Re:Deja Vu. (Score:5, Insightful)

            by thegarbz ( 1787294 ) on Wednesday October 26, 2022 @03:47AM (#62998831)

            "Back" was ambiguous. I was talking about the side opposite to the display connectors, the end that points to the front of the PC. Given the lack of 3.5" bays on modern PCs and the fact that GPUs stopped being insanely long after about the 10x0 series back half a decade ago so we should have ample space on that side.

            • Several of the cards terminate before reaching the edge of the heatsink/fan assembly on that side of the card. Nvidia still put the connector there for many of their rtx2000 cards, but this complicated disassembly with extension cables and glue to ensure they didn't come unplugged.

              The standards for graphic card mounting are so old and outdated that there are problems no matter how things are manipulated. I'm not sure I buy the critiques about bends near the connector either, as it appears that half of the

            • I have a ninety degree board so the "back" points directly at the case fans. The NVIDIA solution, if I had one would point at the side-panel.

          • But we are talking specifically about the 4090 which are huge cards. It is not likely that anything is above or below the card as it takes 4 slots. Putting the connector on the end is a possibility but would require some engineering work. A more likely scenario is Nvidia favored the sleek look of their card over practical considerations.
        • Re:Deja Vu. (Score:5, Funny)

          by arglebargle_xiv ( 2212710 ) on Wednesday October 26, 2022 @02:56AM (#62998775)
          Why on earth do we insist on inventing a neverending series of new incompatible connector for each line of graphics cards? Just slap an IEC 60309 connector on it and you'll be set for the future. For example fitting a 125/100 3P+N+E socket should leave plenty of headroom for future GPU power requirements.
          • The new connector wasn't about GPU power requirements, it was about PSU signaling what kind of power it can make available and what kind of cable is connected so that the GPU knows what power budget it has for boosting.

            Your welding plug (I know you used is as a joke but I feel the need to point out a flaw not related to the joke about size) doesn't have sense pins or other communication so it doesn't fit the bill.

            • Well, in theory it does since it doesn't need any sense pin as the only possible answer to the request "what power level can you supply?" is "yes".
              • That's not really true. In things that draw power from mains, we've solved the "signaling" by making sure that the plugs don't fit if you have an appliance that draws more than can be delivered. However, there are all kinds of dangerous "adapters" out there that allow the truly determined to create hazards.
              • For welding there is only one possible question. For processors and GPUs there are many possible questions, and if the outcome is wrong the possible answers range from "your CPU/GPU cannot distinguish between low- and high-level logic states and has faulted" to "you now have a melted puddle of silicon where the CPU/GPU used to be." Proper CPU/GPU operation depends on a very tight feedback loop between CPU/GPU and power controller.
        • Why on earth we insist and putting the connector in a position where such a bend is required is beyond me. GPUs are big enough that in many cases there's no option other than breaching this bend radius. It would make more sense to have the connector come with right/angle connectors or be in a different location on the card (such as the back, which used to be a problem when our PC cases were full of HDD and GPUs were insanely long, but they've gotten shorter and fatter).

          "That the way it's been done for billions of years." (Slides a $9 billion brand new Gerald R. Ford-class aircraft carrier into a rowboat berth.)

    • The worst thing to me about that is the statement that failure is not unusual after 40 plug/unplug cycles. That's insane. As GN states, most connectors in PCs are good for hundreds if not thousands of plug/unplug cycles. The vendor might be able to get away with saying that the cables should not be bent or stressed, but I don't think a 40-cycle limit is going to hold up. Based on this, it seems pretty likely that these connectors will need to be redesigned or supplemented with some current-balancing pro
  • It isn't called that for nothing.
  • These things already draw 600W peak, TDPs these days are for a âregularâ(TM) workout, crunching numbers will put a significant strain above the designed spec. Just a light overclock and they are drawing 150W above TDP already, combine that with a very thin connector that is often poorly installed or modified for âprettyâ(TM) cable management and water cooling, these things will start smoking a lot more often.

    We have some high power cards with passive cooling in datacenters and workstatio

    • Change it to Anderson SB50. That should be good enough.

      Or redesign the card to take higher voltage, 12V isn't really suitable for those powers.
      400V 3-phase directly to the card should be good for the forseeable future.

      • Re: (Score:2, Informative)

        by Anonymous Coward
        You laugh but bumping the 12V power supply requirement to 48V or even just 24V would significantly reduce the current draw allowing you to scale back on wire thickness and the pin size in plugs and sockets.
  • by gweihir ( 88907 ) on Tuesday October 25, 2022 @08:44PM (#62998255)

    These are regular connectors in a 16 pin housing. Each pin can take 8A without problem. That means the connector can take 768W if everything is perfectly balanced. Call it around 600W real-world. This card only pulls 450W regular. Of course if you plug this connector or the PSU-side in wrong, contaminate it or the pins or do stupid things to your card that causes power-spikes, this connector may well melt. But that is on the user, not the design.

    • by sinij ( 911942 )
      I kept thinking how to exceed even conservatively (i.e., cheap Chinese crap) defined max rating without triggering OCP and couldn't think of anything. I think these are made out of polycarbonate, it melts at 280C. I think temperature rating at max current for this type of insulation would be 75C. As such, it needs to exceed max by quite a bit to melt the connector.
      • by gweihir ( 88907 )

        Indeed. Unless you do stupid crap like plugging in only one of the connectors on the other side. Or plugging them in halfway.

    • There is a difference between a current rating and continuous current rating within an enclosure. The pins rated for 8A will be fine at 8A so long as they are not adjacent other pins and have access to airflow. Group all 16 pins together and place them within a plastic shroud and bad things can happen.

      Derating the current capacity of a conductor is often required when wiring in order to conform to electrical codes. Are you placing the cables within a conduit? Are the cables bundled with other cables?

      • Re: (Score:2, Redundant)

        by willy_me ( 212994 )

        I should also point out that the main form of cooling for these connectors will be the PCB and connected power wires. Heat from the connector will go up the power lines and dissipate to the environment - basically the wire acts as a heatsink. So if you tidy the power lines and perhaps group them together within a cable sheath, you are going to get less cooling.

        At the end of the day there will have been many factors at play. The problem is that the design does not allow for these factors.

      • by gweihir ( 88907 )

        These pins are rated for 8A when installed into these connectors double-row with others adjacent and no airflow, just a maximum ambient temperature. Seriously. Using them singularly, they probably would be rated higher but that is not the way this type of pins get specified.

    • by Z80a ( 971949 )

      According to some people that been in the business like that jay two cents, the problem is most likely how easy it is to bend the pins when trying to install the card.
      The connector requires a clearance of 35mm from the first bend, and that is very easy to get wrong.
      And as bad contacts heat up, you get fire

      • Re:Not plausible (Score:4, Interesting)

        by gweihir ( 88907 ) on Tuesday October 25, 2022 @10:10PM (#62998445)

        According to some people that been in the business like that jay two cents, the problem is most likely how easy it is to bend the pins when trying to install the card.

        That would do it. These connectors have two contact points in the spring side. If you bend a pin, you get only one good contact point or maybe even none. Do that to several contacts and it heats up too much.

        The connector requires a clearance of 35mm from the first bend, and that is very easy to get wrong.

        This may be a requirement the adapter manufacturer has put in there, but the connectors only specify some maximum mechanical load as far as I recall (did read several datasheets for them a few months back because I was doing a custom connector for). But yes, too much mechanical load on the contacts can also get you bad contact and heating up and these 35mm may be what the wires used give you for the maximum force. For example, if somebody bends the whole wire bundle close to the connector by 180 degrees, that may well do it.

        I think the bottom line is that loading a connector of this type up to close to its maximum load is something you can do if only experts and careful people handle it. Throw in a few fat-finger types and you get the observed effects.

        • by AmiMoJo ( 196126 )

          It doesn't help that PSU manufacturers don't supply right angle connectors to reduce the strain.

          I imagine the manufacturers are the ones most concerned about this. If the connector melts they are the ones who will be getting the PSU RMAed, and possibly the graphics card too if that gets damaged. Apparently even at the RRP they can't make much money on a 4090, and some may be losing money on what is considered a halo product.

          • by gweihir ( 88907 )

            Indeed. I guess manufacturers should stop to make these "bleeding edge" stuff. It just does not make much sense. Or maybe the PC buyers are too irrational and will buy something smaller only of the manufacturer also makes the larger thing.

    • Re:Not plausible (Score:4, Informative)

      by Xenx ( 2211586 ) on Tuesday October 25, 2022 @10:08PM (#62998439)

      These are regular connectors in a 16 pin housing.

      It's a 16pin connector. However, 4 of them are for sideband. That leaves 12 pins. The spec indicates 6 pins at 9.2A for a total of 55.2A and 662.4W.

      As to what is happening, so far testing indicates it's a matter of bending the wires too close to the connector. It's causing stress, and imperfect contact, within the connector.

    • Re:Not plausible (Score:5, Informative)

      by DamnOregonian ( 963763 ) on Wednesday October 26, 2022 @01:14AM (#62998673)

      These are regular connectors in a 16 pin housing.

      No. 12 power pins, 4 side-band (not current carrying).

      Each pin can take 8A without problem.

      9.2A.

      That means the connector can take 768W if everything is perfectly balanced.

      660W.

    • Re:Not plausible (Score:4, Interesting)

      by willy_me ( 212994 ) on Wednesday October 26, 2022 @01:50AM (#62998719)

      It should be noted that the IEC only provides a safety rating of 5A. This would be a worst case scenario - so all circuits loaded and in a warm environment. Looks like we have a worst case scenario here.

      With a 5A rating we only have 30A total (there are only 12 pins that carry power), or 360W assuming the power supply maintains 12V output. Below the 450W power draw of the card. Good luck getting IEC certification....

      I speculate that we will eventually have a recall on these boards due to being a fire hazard. Over time, the resistance of the connector will increase due to corrosion. Looks like the card and power supplies use tinned pins - (should have used gold). As the resistance increases the heat will build up and eventually things will melt / combust. Even if this only happens on a small number of boards, it still warrants a recall. And considering that it has already happened and that it is going to get worse with time - expect either a recall or a new firmware that castrates the card.

      And fyi, power-spikes are a non-issue. The issue is with heat which is based on average current level. And you can not plug it in wrong as it is keyed. But contaminating it or the pins is a good example of what could happen.

      • by psmears ( 629712 )

        And you can not plug it in wrong as it is keyed.

        I fear you may be underestimating people's incredible ability to get things wrong ;-) (I guess one possibility is that the plug is somehow not fully engaged - so that the pins are only making partial contact, leading to higher resistance?)

    • by Junta ( 36770 )

      But that is on the user, not the design.

      Frequently not. The design should take into account reality. If you know that the cable will tend to be heading for a 90 degree turn right behind the connector, but your design is susceptible to poor contact or over stress the connector when that happens, that's a design problem. If you release something with a specified cycle of 10 reconnects or so, then you almost certainly have a design issue. If typical household environments have a high likelihood of too much contamination, then your design would be

      • by gweihir ( 88907 )

        Remember that most of these do not melt. Something wrong has to be done to them. But yes if you allow user-installation on something like this, you have to put in generous safety margins. Loading a 600W connector to 600W in that scenario is obviously asking for trouble.

        • by Junta ( 36770 )

          True, this is going off of anecdotes with poor discipline with respect to overall data collection.

          However it jives with what I have heard said from electrical engineers I know when they first saw the specifications. That style connector with that sort of power delivery sounds like it is just asking for trouble, particularly when oriented to head straight for where a case cover usually sits a few centimeters away. Since I heard people call it months ago at work and now precisely their predictions are coming

          • by gweihir ( 88907 )

            However it jives with what I have heard said from electrical engineers I know when they first saw the specifications. That style connector with that sort of power delivery sounds like it is just asking for trouble, particularly when oriented to head straight for where a case cover usually sits a few centimeters away. Since I heard people call it months ago at work and now precisely their predictions are coming to pass in the real world, I'm inclined to give them credit.

            You should give these EEs credit. I did not see it back then, but it is not much of a surprise. Sure, this is not asking for every single one of these to melt, but it was pretty clear it would be more than a few and given the price-tag on the cards, that is just completely unacceptable. Connectors melting should be _rare_, like 1 in 1'000'000 or so when new. That is why you either restrict the load to something like 50% of rated load or make them extra sturdy. Just like in all other engineering, reliability

  • It sounds like the wiring needs to be upgraded to handle the current. That would go along with a better grade of plastic maybe PPS (Polyethelene Sulfide Plastic) which is rated up to 212C but that would require some data other than "it melted." It would be great to use an IR thermometer to see how high the temperatures go and validate the wire grade and temperature rating of that as well..

    • by gweihir ( 88907 )

      No, it is the contact pins loaded up to the limit in a relatively cheap design (Molex Mini-Fit Jr.). Any problem with the installation, like mechanical side-load on the connector, and the connector does not have reserves and heats up. Smart electrical designers always put in reserves. Guess the ones doing this design were not smart and thought you can go up to the maximum rated load with no problems.

    • It sounds like the wiring needs to be upgraded to handle the current.

      The wires aren't melting. The connector is.

  • Seems like an extreme use case in both cases as both are sold as OC cards. Maybe run the heatsink up into space?
  • by SciCom Luke ( 2739317 ) on Wednesday October 26, 2022 @01:46AM (#62998717)
    Sure... The go for 2600 euro. And that difference is not only taxation.
    This card is clearly not designed for customers who give the tiniest shit about their electricity bill.
  • RTX 4090 sucks 450W down though the compact power connector, but factory overclocked models can push that up to around 500W, and manual overclocking could increase that to 600W or more. The Redditor asserts he was simply playing Red Dead Redemption 2, with a relatively modest GPU load of 400W, when the cable melted, which appears to rule out extreme overclocking.
  • A slashdot user is throwing shade at Reddit. Hilarious.

"All the people are so happy now, their heads are caving in. I'm glad they are a snowman with protective rubber skin" -- They Might Be Giants

Working...