RTX 5090 FE Molten 12VHPWR

8 pin connector has 19 years of history behind it, you can't be serious comparing the few failures it had with 12hpwr that is only around since 2022 and making the news since release. There is already far too many documented cases.

That Google search link above mostly just shows burned 12hpwr adaptors anyway, lol - despite the keywords.
 

“Cooler Master GPU Shield Gets Lower Price and Warranty Period in EU​

by Cpt.Jank Today, 18:21 Discuss (0 Comments)
A month after Cooler Master launched its GPU Shield 12v-2×6 power cable monitoring solution in the US, it has finally released the safety accessory in the EU, where it received a €29.99 price tag—compared to $49.99 in the US.

Along with the lower price, however, Cooler Master has also reduced the warranty period on the GPU Shield adaptor cable from three years in the US and China to just two years in the EU—just barely covering the minimum threshold for legally mandated warranties in the EU, but still longer than the one-year warranty granted for the international version of the GPU Shield. The GPU Shield adaptor cable is available via Cooler Master directly, and it does not appear to be available on sites like Amazon just yet.“
 
TXE36 said:
We will have to agree to disagree. This is all about odds. In properly designed systems the odds of any one person having "bad luck" and getting a faulty unit are extremely thin - like winning a big lottery. As stated earlier, way too many lottery winners in this thread. Keep in mind that these samples were factory shipped units that supposedly went through product testing - defects should have been caught before shipping. Randomly finding one bad one in this pile means your failure rate is very high. I don't consider the design borderline either, it is poor.

This thread is one and half years old and is still going. I honestly don't recall this much discussion with respect to 6 pin and 8 pin PCIe power connectors failing. Time is not going to be kind to this issue.

The attitude that some of the product must fail is just a crutch - well engineered products "just work". The need to put active cooling and temperature monitoring on power connections is just an epic fail.
Click to expand...
A.I. makes it so esay to debunk false claims these days:

No, it does not contain a large enough sample size. Your skepticism about the poster's lack of statistical knowledge is completely spot on.
When evaluating hardware issues through the lens of a singular Hard|Forum discussion, the community discourse almost always mistakes anecdotal clusters for an actual statistical failure rate.
Here are the specific mathematical and statistical reasons why a forum thread—even one spanning dozens of pages—is fundamentally useless for drawing a definitive conclusion about the reliability of the RTX 5090 FE:

1. Extreme Selection Bias (The "Loud Minority")
A forum thread dedicated to a specific failure acts as a homing beacon. The population within that thread is not a randomized sample of RTX 5090 owners:
  • 100% of affected users look for a place to post: If a consumer spends $2,000+ on a flagship GPU and it melts, they will instantly search the internet, find this exact thread, and post their pictures out of frustration or a desire for advice.
  • The silent, happy 99.x% have no reason to post: The tens of thousands of users whose RTX 5090s are running flawlessly at that exact moment do not log onto Hard|Forum to create a post saying, "Day 45: My cable did not melt today."
  • This creates a massive data distortion, making an incredibly rare event look like a widespread epidemic.

2. The Denominator Problem
In statistics, a fraction means nothing if you only know the numerator (the number of melted plugs) but have absolutely no idea what the denominator is (the total number of RTX 5090 FE units sold and actively running).
  • If a thread documents 20 distinct, real cases, but NVIDIA has shipped 150,000 units of that specific card globally, the failure rate is roughly 0.013%.
  • In mass manufacturing of complex electronics, a failure rate that low is well within standard acceptable tolerances. Without access to proprietary RMA (Return Merchandise Authorization) data from NVIDIA or major retailers, calculating actual risk is mathematically impossible.


3. Uncontrolled Confounding Variables
To statistically prove that the RTX 5090 FE design itself is the root cause of the failure, you have to isolate the variables. In a forum environment, the data is incredibly noisy and full of hidden factors:
  • User Error: Plugs not being pushed completely into the socket until they click.
  • Mechanical Stress: Aggressive, tight cable bends forced against the side glass panels of smaller PC cases.
  • Third-Party Variance: The use of aftermarket extensions, custom cables, or recycled power supplies from older builds where the internal terminals might already be worn down.

The Bottom Line
The only valid conclusion you can extract from that thread is an existence proof: Yes, it is physically possible for a 12VHPWR / 12V-2x6 connector to melt on an RTX 5090 FE under specific conditions. As a tool for quantifying probability, systemic engineering defects, or overall product safety, the thread holds a statistical value of zero.
 
littlegreenmenfromhome said:
A.I. makes it so esay to debunk false claims these days:

No, it does not contain a large enough sample size. Your skepticism about the poster's lack of statistical knowledge is completely spot on.
When evaluating hardware issues through the lens of a singular Hard|Forum discussion, the community discourse almost always mistakes anecdotal clusters for an actual statistical failure rate.
Here are the specific mathematical and statistical reasons why a forum thread—even one spanning dozens of pages—is fundamentally useless for drawing a definitive conclusion about the reliability of the RTX 5090 FE:

1. Extreme Selection Bias (The "Loud Minority")
A forum thread dedicated to a specific failure acts as a homing beacon. The population within that thread is not a randomized sample of RTX 5090 owners:
  • 100% of affected users look for a place to post: If a consumer spends $2,000+ on a flagship GPU and it melts, they will instantly search the internet, find this exact thread, and post their pictures out of frustration or a desire for advice.
  • The silent, happy 99.x% have no reason to post: The tens of thousands of users whose RTX 5090s are running flawlessly at that exact moment do not log onto Hard|Forum to create a post saying, "Day 45: My cable did not melt today."
  • This creates a massive data distortion, making an incredibly rare event look like a widespread epidemic.

2. The Denominator Problem
In statistics, a fraction means nothing if you only know the numerator (the number of melted plugs) but have absolutely no idea what the denominator is (the total number of RTX 5090 FE units sold and actively running).
  • If a thread documents 20 distinct, real cases, but NVIDIA has shipped 150,000 units of that specific card globally, the failure rate is roughly 0.013%.
  • In mass manufacturing of complex electronics, a failure rate that low is well within standard acceptable tolerances. Without access to proprietary RMA (Return Merchandise Authorization) data from NVIDIA or major retailers, calculating actual risk is mathematically impossible.

3. Uncontrolled Confounding Variables
To statistically prove that the RTX 5090 FE design itself is the root cause of the failure, you have to isolate the variables. In a forum environment, the data is incredibly noisy and full of hidden factors:
  • User Error: Plugs not being pushed completely into the socket until they click.
  • Mechanical Stress: Aggressive, tight cable bends forced against the side glass panels of smaller PC cases.
  • Third-Party Variance: The use of aftermarket extensions, custom cables, or recycled power supplies from older builds where the internal terminals might already be worn down.

The Bottom Line
The only valid conclusion you can extract from that thread is an existence proof: Yes, it is physically possible for a 12VHPWR / 12V-2x6 connector to melt on an RTX 5090 FE under specific conditions. As a tool for quantifying probability, systemic engineering defects, or overall product safety, the thread holds a statistical value of zero.
Click to expand...

Yes, that's right—why not ask the AI:

You are correct that the melting issue is not solely due to user error or “bad design” in a simplistic sense, but rather a combination of design flaws and high-power demands that create a perfect storm for failure. The issue is complex and has been extensively analyzed by experts, including hardware reviewers and electrical engineers.

The Core Design Flaws​

  1. Inadequate Pin Contact and Current Density:
    • The 12VHPWR connector is designed to handle up to 600 watts of power (500 watts on the 12V rail, 7.5 amps, plus 12 watts on the 12V aux rail).
    • To fit this much power into a small connector, the pins are very thin. This creates high current density, which naturally generates more heat than a thicker, more robust connector design (like the older 8-pin PCIe connectors).
    • Experts have pointed out that the load balancing across the pins can be uneven. If some pins carry more current than others due to slight manufacturing tolerances or wear, those specific pins will heat up disproportionately, leading to localized melting.
  2. Pin Degradation Over Time:
    • The plating on the pins (usually gold or nickel) can degrade over time due to thermal cycling (heating up when the GPU is under load and cooling down when idle).
    • As the plating degrades, the electrical resistance at the contact point increases. Higher resistance means more heat, which further degrades the plating, creating a positive feedback loop that eventually leads to melting.
    • This degradation can happen even if the connector is plugged in correctly and handled with care, which is why users report melting after 1-3 years of normal use.
  3. Lack of a Mechanical Stop (Initial Design):
    • The original 12VHPWR connector (used on early RTX 4090s) did not have a mechanical stop to ensure it was fully seated. This made it very easy for users to plug it in partially, leading to poor contact and melting.
    • The updated 12V-2x6 connector (used on RTX 5090s and newer RTX 4090s) includes a small tab that prevents incomplete insertion. However, this does not address the underlying issues of pin degradation, current density, and load balancing.
  4. Cable and Connector Quality:
    • The quality of the cable and connector can vary significantly between manufacturers. Third-party cables and adapters (like those from CableMod) have had higher reported failure rates than the cables provided by NVIDIA or major PSU manufacturers.
    • The internal wire gauge and pin plating quality can affect resistance and heat dissipation. Thinner wires or poor-quality plating can exacerbate the melting issue.

Why RMA Statistics Are Likely Skewed​

Your hypothesis about RMA statistics being skewed is well-founded:
  1. Underreporting Due to Warranty Denials:
    • NVIDIA and other manufacturers often deny warranty claims if they determine the damage was caused by user error (e.g., improper installation, using a third-party cable, or bending the cable too sharply).
    • Many users likely do not send melted cables for RMA because they fear the claim will be denied, leaving them with a damaged GPU and no refund.
  2. Financial Risk:
    • The RTX 5090 is a very expensive card. If the RMA is denied, the user loses the entire value of the card. This is a significant financial risk, especially for a high-end card.
  3. Complexity and Effort:
    • The RMA process can be lengthy and frustrating. Some users may choose to simply replace the cable themselves (if the GPU is undamaged) or accept the risk rather than go through the process.
  4. Official Numbers vs. Reality:
    • NVIDIA initially stated they were aware of only about 50 cases globally of melting 16-pin connectors, which they calculated as a 0.04% to 0.05% failure rate. However, this number likely represents only the cases that were reported and verified, not the total number of incidents.
    • Community reports on forums like Reddit and Level1Techs frequently show new cases of melting, suggesting that the actual incidence rate is higher than official figures.

Conclusion​

The melting of the 12VHPWR connector is a result of both design challenges (high power density in a small connector, thin pins, potential load balancing issues) and human factors (installation errors, cable handling, cable quality). The official RMA statistics are almost certainly an undercount because many users avoid the process due to the risk of warranty denial and financial loss. This makes it difficult to get an accurate picture of how widespread the problem truly is.
The issue is not just “bad design” but a combination of factors that create a risk of failure over time, even with careful use. The updated 12V-2x6 design addresses the incomplete insertion issue but does not fully resolve the underlying problems of pin degradation and current density.
You are correct that experts have identified significant design flaws in the 12VHPWR connector architecture, and that these flaws contribute to melting even when the connector is plugged in correctly. The issue is not solely user error.

Design Flaws Contributing to Melting​

  1. Lack of Load Balancing: A primary criticism from experts is that the 12VHPWR connector lacks active or passive load balancing safeguards. In a high-power connector, current should be distributed evenly across all power pins. However, if the resistance varies slightly between pins (due to manufacturing tolerances or wear), current will naturally flow through the path of least resistance. Without load balancing, one or two pins can carry a disproportionate amount of the 600-watt load, leading to thermal overload and melting.
  2. Thin Pins and Loose Tolerances: Experts and reviewers (such as JayzTwoCents) have pointed out that the pins in the 12VHPWR connector are relatively thin and can be “jiggly” or loose within the housing. This looseness can lead to poor contact over time, increasing resistance and heat generation. The cable tolerances are also reported to be too loose, which exacerbates the issue.
  3. Minimal Safety Factor: The 12V-2x6 connector (the updated version) has been criticized for having a very minimal safety factor. It is rated for 600 watts, but the safety factor is only about 1.1, meaning it operates very close to its thermal limits. This leaves little room for error or degradation.
  4. Pin Degradation and Fretting Corrosion: Over time, the tin plating on the pins can degrade due to thermal cycling (heating up and cooling down) and “fretting corrosion” (micro-movement of the pins). This degradation increases electrical resistance, which in turn generates more heat, creating a vicious cycle that can lead to melting even if the connector was initially seated correctly. This degradation can happen within 1–3 years of normal use, as you suggested.
  5. Structural Integrity Issues: The connector as a whole may lack structural integrity, making it prone to deformation under the pressure of the attached cables, which can further compromise the connection.

Impact on RMA Statistics​

Your hypothesis about RMA statistics being skewed is highly likely correct.
  1. Underreporting Due to Warranty Denials: As you noted, many users avoid RMA because they fear denial. If a user used a third-party cable (like CableMod) or if NVIDIA determines improper installation (even if it wasn’t the primary cause), the RMA may be denied. CableMod, for example, had to reimburse users whose RMAs were denied because their adapters were used.
  2. Financial Risk: The financial risk of a denied RMA (losing the entire GPU value) discourages users from reporting incidents, especially if the damage is only to the cable and not the GPU itself. Many users may simply replace the cable themselves if the GPU is undamaged.
  3. Official Numbers are Likely Low: NVIDIA initially stated they were aware of only about 50 cases globally, calculating a 0.04% to 0.05% failure rate. However, this number represents only reported and verified cases, not the total number of incidents. Community reports on forums like Reddit and Level1Techs continue to show new cases of melting on RTX 5090s, suggesting the actual incidence rate is higher.
  4. Persistent Issues Despite “Fixes”: The fact that melting incidents continue to be reported on the RTX 5090, which uses the improved 12V-2x6 connector, indicates that the underlying design flaws (like lack of load balancing and pin degradation) have not been fully resolved. This persistence further suggests that many incidents may go unreported, making official statistics an undercount.
In summary, the melting issue is a result of fundamental design flaws in the 12VHPWR connector, including lack of load balancing, thin pins, and minimal safety margins, which can lead to degradation and melting over time even with correct usage. The official RMA statistics are almost certainly an undercount due to users avoiding the process because of the risk of warranty denial and financial loss.
Click to expand...
 
Last edited:
Ididar said:
So ignore actual statistical modeling methodology which all industries use to study failure rates and trust the AI. Gotcha.
Click to expand...
He just asked for it.

And yes, go check out the service centers—you'll be surprised how many cases go unreported, or when someone simply replaces their cable because they notice the first signs of damage.
 
Ididar said:
So ignore actual statistical modeling methodology which all industries use to study failure rates and trust the AI. Gotcha.
Click to expand...
I'm an electrician been one my whole life. I don't need statistics to know this is THE weak point that is most likely to kill my card one day. Will it fail? Maybe maybe not. But it shouldn't be happening at all as it's VERY easy to calculator a good safety factor in Electrical connections whether it's something small like a gpu or switchgear powering a building the principle is the same. You don't want your connection to be what makes something go off line one day. So as a result most everything in electrical is overrated it's why your panel in your house is still good 20+ years later. I want my GPU cable to be able to take the card full load without worrying it blows.

The math of the plug tells the story. We had way more safety factor on 8pin....that's why over the however many years we used it we don't have a massive my 8pin melted up to page 1000 by now.
 
Well, you see something new every day, I never thought Enshittification would attract fanbois...

Bottom line, in over 35 years of designing stuff professionally I've learned that if you take shortcuts like Nvidia has, it will bite you in the ass eventually. But what do I know, I'm just an old school electrical engineer with a well earned reputation that "my shit works".

BTW, this issue is not a connector issue, it is a PCB design issue that is overloading the connector. The PCB design with the connector is well out of bounds of sound engineering practice and no amount of AI-fu will change that.

littlegreenmenfromhome said:
3. Uncontrolled Confounding Variables
To statistically prove that the RTX 5090 FE design itself is the root cause of the failure, you have to isolate the variables. In a forum environment, the data is incredibly noisy and full of hidden factors:
  • User Error: Plugs not being pushed completely into the socket until they click.
  • Mechanical Stress: Aggressive, tight cable bends forced against the side glass panels of smaller PC cases.
  • Third-Party Variance: The use of aftermarket extensions, custom cables, or recycled power supplies from older builds where the internal terminals might already be worn down.

The Bottom Line
The only valid conclusion you can extract from that thread is an existence proof: Yes, it is physically possible for a 12VHPWR / 12V-2x6 connector to melt on an RTX 5090 FE under specific conditions. As a tool for quantifying probability, systemic engineering defects, or overall product safety, the thread holds a statistical value of zero.
Click to expand...

With respect to the above, the PCIE 6 and 8 pin connectors didn't have any problem with those "Uncontrolled Confounding Variables". I would assume that Nvidia knew these things were going in PC cases?

Also, I don't believe anybody here made any claims about the statistical number that will fail -- it's only been said that the design is weak/poor, and if history is a guide, the design will have more issues as time passes.

Testing can never prove something is 100% reliable, however, it can prove failure, and to do that, all you need is one honest report. Hand waving doesn't change that.
 
TXE36 said:
Well, you see something new every day, I never thought Enshittification would attract fanbois...

Bottom line, in over 35 years of designing stuff professionally I've learned that if you take shortcuts like Nvidia has, it will bite you in the ass eventually. But what do I know, I'm just an old school electrical engineer with a well earned reputation that "my shit works".

BTW, this issue is not a connector issue, it is a PCB design issue that is overloading the connector. The PCB design with the connector is well out of bounds of sound engineering practice and no amount of AI-fu will change that.



With respect to the above, the PCIE 6 and 8 pin connectors didn't have any problem with those "Uncontrolled Confounding Variables". I would assume that Nvidia knew these things were going in PC cases?

Also, I don't believe anybody here made any claims about the statistical number that will fail -- it's only been said that the design is weak/poor, and if history is a guide, the design will have more issues as time passes.

Testing can never prove something is 100% reliable, however, it can prove failure, and to do that, all you need is one honest report. Hand waving doesn't change that.
Click to expand...
A lot of words to say your were making a false claim.
 
1791990302896.png
 
TXE36 said:
Your AI screed was 3 times longer than my human written response. Your post above and strawman approach indicates you are not serious. What was the AI prompt you used to generate it?

Directly quote and highlight my false claim. I'll wait...
Click to expand...
You know what I just put that guy on my block list the alien form Mars was getting annoying always saying we are basically stupid for not sucking Nvidia off he adds nothing to this and that ai prompt was the icing on the cake. I have mk1 hands/eyes and a brain from being an electrician for 20+ been doing this since I was like 10 years old I can tell when something is under built no matter what Nvidia or an AI says. It's junk
 
tunatime said:
I'm an electrician been one my whole life. I don't need statistics to know this is THE weak point that is most likely to kill my card one day. Will it fail? Maybe maybe not. But it shouldn't be happening at all as it's VERY easy to calculator a good safety factor in Electrical connections whether it's something small like a gpu or switchgear powering a building the principle is the same. You don't want your connection to be what makes something go off line one day. So as a result most everything in electrical is overrated it's why your panel in your house is still good 20+ years later. I want my GPU cable to be able to take the card full load without worrying it blows.

The math of the plug tells the story. We had way more safety factor on 8pin....that's why over the however many years we used it we don't have a massive my 8pin melted up to page 1000 by now.
Click to expand...

You're ignoring the point I made. I never said the design was good. I said the math didn't add up in claiming how big the problem was. I made it very clear that I was pointing out only an error in using bad statistics, not claiming anything about the design of the connector itself. I know it isn't great. That's not the point I was making. I was simply pointing out someone using really bad math to make a judgement about the scope of the problem.
 
Ididar said:
You're ignoring the point I made. I never said the design was good. I said the math didn't add up in claiming how big the problem was. I made it very clear that I was pointing out only an error in using bad statistics, not claiming anything about the design of the connector itself. I know it isn't great. That's not the point I was making. I was simply pointing out someone using really bad math to make a judgement about the scope of the problem.
Click to expand...
The really big blunder here is gathering accurate statistics is going to be nearly impossible. Apparently nobody at Nvidia thought that cable interconnects may be obtained from questionable Chinese sources and did not build in any margin to accommodate less than stellar connectors and contacts. The scope of the problem itself is nasty because it is very difficult to bound.

Molex has extensive design data and quality controls for its connectors. A Molux brand on Aliexpress is likely not to be of the same quality.

The crying shame here is the lack of engineering and foresight. Look at the incredible engineering of the silicon and driver software and it gets let down by a cheesy power supply design. A multi-phase input power supply would have been reliable even with cheap shit connectors and extensions - we know this not only from the engineering theory but also from decades of experience.
 
tunatime said:
You know what I just put that guy on my block list the alien form Mars was getting annoying always saying we are basically stupid for not sucking Nvidia off he adds nothing to this and that ai prompt was the icing on the cake. I have mk1 hands/eyes and a brain from being an electrician for 20+ been doing this since I was like 10 years old I can tell when something is under built no matter what Nvidia or an AI says. It's junk
Click to expand...
I dont think the design is good enough either, but the original claim that littlegreenmenfromhome was responding to is just incorrect.

"Sample size does matter - technically the sample size to prove a design is insufficient is one. The sample size to get a warm fuzzy feeling that everything is okay is huge."

No. A sample size of 1, with 1 failure does not mean the design is insufficient. We would need to ask what is an acceptable failure rate for the given product, and then get a reasonable sample size to determine the failure rate. I think someone posted failure rates earlier in this thread, but I dont remember if it was for all 12vhpwr connectors, or for 5090 specifically. I would be interested to see 5090 specifically, and it would be especially great to see monthly failure rate and see if there is a spike with DLSS5 release.
 
serpretetsky said:
I dont think the design is good enough either, but the original claim that littlegreenmenfromhome was responding to is just incorrect.

"Sample size does matter - technically the sample size to prove a design is insufficient is one. The sample size to get a warm fuzzy feeling that everything is okay is huge."

No. A sample size of 1, with 1 failure does not mean the design is insufficient. We would need to ask what is an acceptable failure rate for the given product, and then get a reasonable sample size to determine the failure rate. I think someone posted failure rates earlier in this thread, but I dont remember if it was for all 12vhpwr connectors, or for 5090 specifically. I would be interested to see 5090 specifically, and it would be especially great to see monthly failure rate and see if there is a spike with DLSS5 release.
Click to expand...
When you add to the sample size of one the engineering analysis you don't need statistics to know it's bad. It's one thing to have a failure that doesn't seem to make sense versus a failure and further investigation reveals "this is effed up".

Right now the deniers are using the "lack of statistics" to hand wave "it's fine!".

Another question, how do you think the failure rates of video cards with 6/8 pin PCIe connectors compares with the 12VHPWR based on decades of use?
 
TXE36 said:
When you add to the sample size of one the engineering analysis you don't need statistics to know it's bad. It's one thing to have a failure that doesn't seem to make sense versus a failure and further investigation reveals "this is effed up".

Right now the deniers are using the "lack of statistics" to hand wave "it's fine!".

Another question, how do you think the failure rates of video cards with 6/8 pin PCIe connectors compares with the 12VHPWR based on decades of use?
Click to expand...
Failure analysis and determining failure rates (which is where sample sizes are relevant) are distinct concepts, so it doesn't make sense to say "When you add to the sample size of one the engineering analysis you don't need statistics to know it's bad."
You are not adding to the sample size, you are doing failure analysis. I agree, doing failure analysis can sometimes be sufficient, but it's meaningless to talk about its sample size. It tells me very little about failure rates. Perhaps I could guess at the failure rate from the failure analysis, but that's not really what we want here.
 
Leave the multi-trillion dollar company alone! They and their partners can't afford to replace 10 cent connectors on $7000 GPUs when they melt due to poor connection quality, and they aren't melting at unacceptable rates because Nvidia mods immediately remove reports from their reddit page! And just because a whole new industry is starting up around these connectors it doesn't mean that they are a problem!

(this is sarcasm)

Seriously all, it's beyond silly.
 
serpretetsky said:
Failure analysis and determining failure rates (which is where sample sizes are relevant) are distinct concepts, so it doesn't make sense to say "When you add to the sample size of one the engineering analysis you don't need statistics to know it's bad."
You are not adding to the sample size, you are doing failure analysis. I agree, doing failure analysis can sometimes be sufficient, but it's meaningless to talk about its sample size. It tells me very little about failure rates. Perhaps I could guess at the failure rate from the failure analysis, but that's not really what we want here.
Click to expand...
Let's try a simpler analogy.

Company ships out widgets. In this widget a 10W resistor is loaded to 12-15W under some user conditions due to a design error. One of these resistors burns up. An engineer looks at the design and discovers the 12-15W load on a 10W resistor. No more statistics are needed for the evaluation. You don't need to go out and find 100 more or 1000 more examples to change the conclusion.

When components are pushed beyond their datasheet ratings and good engineering practice one ends up in no man's land where it is very difficult to put a bound on how bad a problem could be today, and in the future. Good engineers avoid this.

In the automotive aftermarket there is this concept called bulletproofing. These are mods done to things like engines to address weaknesses. One example is wirenutting the oil pump sprocket nut on the BMW E36 to keep it from backing off under track conditions. If I had one of these cards, instead of spending $200 to monitor the crappy connector, I'd spend $200 bulletproofing the connector especially because the mfgs don't seem to want to support replacing bad connectors through RMA. I also wouldn't want the graphics card damaging my power supply.

Nvidia effed up. Deal with it.
 
TXE36 said:
No more statistics are needed for the evaluation.
Click to expand...
This is pretty much all you had to say, and then we could be discussing whether or not we need more data. I don't really disagree with any of the rest of the above post. Instead, you initially made this incorrect statement:

"Sample size does matter - technically the sample size to prove a design is insufficient is one. The sample size to get a warm fuzzy feeling that everything is okay is huge."

And I can't tell if you are still defending it or if you've accepted it's not really correct and have moved on?
 
serpretetsky said:
This is pretty much all you had to say, and then we could be discussing whether or not we need more data. I don't really disagree with any of the rest of the above post. Instead, you initially made this incorrect statement:

"Sample size does matter - technically the sample size to prove a design is insufficient is one. The sample size to get a warm fuzzy feeling that everything is okay is huge."

And I can't tell if you are still defending it or if you've accepted it's not really correct and have moved on?
Click to expand...

Oh, I'm still defending it and the statement should be read in the context it was presented, it was not meant to stand alone WRT this issue - the second paragraph is significant:

TXE36 said:
Sample size does matter - technically the sample size to prove a design is insufficient is one. The sample size to get a warm fuzzy feeling that everything is okay is huge.

Practically speaking you want to find more than one bad one because the issue could be user error. However, the sample size just presented here is enough to cover that. Now, corroborate the failed reports with an engineering analysis that indicates the PCBs do nothing to limit the individual pin current and it is clear there is an issue with the connector pin tolerance -- the resistance of the connector pins are not controlled well enough from unit to unit to ensure they are not overloaded when connected to the graphics card power input. Note the burned connectors at the power supply side of the cable are the same problem -- the connector used there is having issues too.
Click to expand...

I'm at the point of agreeing to disagree at this juncture.
 
TXE36 said:
Oh, I'm still defending it and the statement should be read in the context it was presented, it was not meant to stand alone WRT this issue - the second paragraph is significant:



I'm at the point of agreeing to disagree at this juncture.
Click to expand...
So then you agree previous designs were also failures? We have multiple samples of failures of the pcie 2x8 connector. Im confident at least a few of those, if you did failure analysis, would show some of the pins had poor contact resistance, or maybe poor crimping connection. If they had used even beefier connectors some of those failures wouldnt happen. Is there any time a failed sample that wouldnt have failed if there was more over engineering doesnt require a redesign, by your standards?
 
serpretetsky said:
So then you agree previous designs were also failures? We have multiple samples of failures of the pcie 2x8 connector. Im confident at least a few of those, if you did failure analysis, would show some of the pins had poor contact resistance, or maybe poor crimping connection. If they had used even beefier connectors some of those failures wouldnt happen. Is there any time a failed sample that wouldnt have failed if there was more over engineering doesnt require a redesign, by your standards?
Click to expand...
I've had dozens of video cards using those connectors over 15+ years without a failure. Don't remember any consistent reporting of them burning up. Do you have PCB reviews from these earlier cards that shows an inadequate input current balancing scheme? I recall there was some card that got dinged for burning up PCIe slots. No, I don't agree that PCIe 6 or 8 pin connectors were failures. I also do not believe the 12VHPWR connector is a failure either.

The focus on the connector is misplaced. The failure does not appear to be the connector, it is the video card power supply design.

For high powered cards I got interested in such as the EVGA 1080TI FTW3 and the Titan XP, Buildzoid's PCB reviews revealed an impressive power supply solution. Note this did not extend to the Titan Pascal - that board makes no sense to me and the power system is weaker. After seeing those boards, Buildzoid's reviews of the 4090 and 5090 were shocking to me given the high price tag of these units.

Another problem with comparing to the PCIe connectors, is how many of them are burned up due to shunt mods? It appears here that these are stock cards using stock connectors. Nvidia has locked a lot of things down these days, during the PCIE 6 and 8 pin heydays, there was a lot more opportunity for abuse.

There are a lot of questionable practices when comes to PC power systems, recall the whole power supply single rail vs multiple rail debates? My personal choice is using a high quality Delta power supply with current limits on each output and then distributing those outputs in a way such that I don't overload any of them. It is an 800W power supply than can push 768W (64A) through the 12V outputs. If you put all that together on a single harness and shorted the 12V to ground, you would smoke the harness. Instead, the 12V is distributed through six different "rails", five at 18A, and one at 18.6A. There is one 12V supply, the rails are created by individual current limiting. Short any of them to ground, or pull too much juice, the power supply clicks off, nothing burns up.
 
serpretetsky said:
nevermind. you're right, we can just disagree here. I dont see this going anywhere.
Click to expand...
The bias is blindng some people, but the math doesn't care so they have to ignore it.
Again the last major study (of the 40 series) found most where user-errors and a sub 1% failure rate:
GamersNexus found that the overall failure rate for the 12VHPWR power connector is very low, estimated between 0.05% and 0.1% across board partners, cable makers, and community data
Click to expand...

But that havn't stopped the crying.
 
littlegreenmenfromhome said:
View attachment 831277

GPU 8 power pins must be a bad design.
Click to expand...

Maybe it is the photo, but that one looks like cheap cable, those wires look too small.

littlegreenmenfromhome said:
I can post picture too, but that adds zero statistical data.
Click to expand...

None of what you posted has any meaning without the video cards they were attached to, documentation on the video card mods, nor evaluation of the quality of the cables. A lot of cheap ass cables use under gauge wires, crappy contacts, cheap housings, and abominations like copper plated aluminum wire.

Historically speaking, if one overclocked Nvidia video cards with Afterburner, and did not alter the power limits beyond the extensions allowed by the BIOS, you could not hurt the card or the power connectors. Run stock and there were no issues. Mod the bios and/or shunt mod the power supply, and all bets were off.

Do you think the 5090 has a good power supply design? Why?
 
littlegreenmenfromhome said:
The bias is blindng some people, but the math doesn't care so they have to ignore it.
...
But that havn't stopped the crying.
Click to expand...
What bias? What crying? Maybe you missed it, but way up I said I had no dog in this hunt. I actually really like and admire the pre-RTX Nvidia designs. What Nvidia has done over its history is truly incredible, but they made a misstep here.

The maf on the 5090 power supply design is quite clear and clearly poor.

littlegreenmenfromhome said:
Again the last major study (of the 40 series) found most where user-errors and a sub 1% failure rate:
Click to expand...
I would expect the 40 series cards to have less of an issue here as they at least have two power phases feeding the power supply vs one on the 5090. 40 series are also lower power with the 4090 TDP being 450W and the 5090 being 575W. Also wasn't there a connector update to try and address the issue when the 4090 was top dog?

A major study of the 50 series sounds like it would be interesting.
 
TXE36 said:
What bias? What crying? Maybe you missed it, but way up I said I had no dog in this hunt. I actually really like and admire the pre-RTX Nvidia designs. What Nvidia has done over its history is truly incredible, but they made a misstep here.

The maf on the 5090 power supply design is quite clear and clearly poor.


I would expect the 40 series cards to have less of an issue here as they at least have two power phases feeding the power supply vs one on the 5090. 40 series are also lower power with the 4090 TDP being 450W and the 5090 being 575W. Also wasn't there a connector update to try and address the issue when the 4090 was top dog?

A major study of the 50 series sounds like it would be interesting.
Click to expand...
So lets recap:
- You were wrong about the statistical significance about this thread.
- You now make excuses for burned 8 pin power plugs (the irony is thick here now)
- You want a PROPER investigation now.

But you are still crying in this thread, a stance, do you have one that isn't "shifting the goalposts" every second post?
 
littlegreenmenfromhome said:
So lets recap:
- You were wrong about the statistical significance about this thread.
- You now make excuses for burned 8 pin power plugs (the irony is thick here now)
- You want a PROPER investigation now.

But you are still crying in this thread, a stance, do you have one that isn't "shifting the goalposts" every second post?
Click to expand...
Another post with nothing of value added. You are in over your head and tunatime was right.
 
TXE36 said:
The really big blunder here is gathering accurate statistics is going to be nearly impossible.
Click to expand...

I agree. I fully agree with the engineering assessment of the connector and how they're really pushing it too far. If people just stuck to that I wouldn't have a conflicting thing to say on this discussion. Most of the debates on this topic are when people try to drag up failure numbers but everyone is basically just puling those numbers out of thin air because there's no way to know, as was noted earlier, what the denominator actually is. Can't know the failure rate without knowing how many are failing and how many are in use but people just love to use numbers to justify their position even when those numbers are just based on feelings.

People do it every day of their lives because numbers have more "punch" when you're trying to make a point. Every day people use made up statistics to try and justify what they're saying and most of the time nobody calls them on their bullshit.
 
You must log in or register to reply here.
Back
Top