Most active commenters
  • ryao(4)
  • lukeschlather(3)

←back to thread

Nvidia won, we all lost

(blog.sebin-nyshkim.net)
977 points todsacerdoti | 26 comments | | HN request time: 1.38s | source | bottom
1. rkagerer ◴[] No.44469388[source]
I am a volunteer firefighter and hold a degree in electrical engineering. The shenanigans with their shunt resistors, and ensuing melting cables, is in my view criminal. Any engineer worth their salt would recognize pushing 600W through a bunch of small cables with no contingency if some of them have failed is just asking for trouble. These assholes are going to set someone's house on fire.

I hope they get hit with a class action lawsuit and are forced to recall and properly fix these products before anyone dies as a result of their shoddy engineering.

replies(4): >>44469514 #>>44469723 #>>44469778 #>>44472796 #
2. rkagerer ◴[] No.44469514[source]
Apparently somebody did sue a couple years back. Anyone know what happened with the Lucas Genova vs. nVidia lawsuit?

EDIT: Plantiff dismissed it. Guessing they settled. Here are the court documents (alternately, shakna's links below include unredacted copies):

https://www.classaction.org/media/plaintiff-v-nvidia-corpora...

https://www.classaction.org/media/plaintiff-v-nvidia-corpora...

A GamersNexus article investigating the matter: https://gamersnexus.net/gpus/12vhpwr-dumpster-fire-investiga...

And a video referenced in the original post, describing how the design changed from one that proactively managed current balancing, to simply bundling all the connections together and hoping for the best: https://youtu.be/kb5YzMoVQyw

replies(3): >>44469586 #>>44469721 #>>44469983 #
3. shakna ◴[] No.44469586[source]
> NOTICE of Voluntary Dismissal With Prejudice by Lucas Genova (Deckant, Neal) (Filed on 3/10/2023) (Entered: 03/10/2023)

Sounds like it was settled out of court.

[0] https://www.docketalarm.com/cases/California_Northern_Distri...

4. middle-aged-man ◴[] No.44469721[source]
Do those mention failing to follow Underwriters Laboratory requirements?

I’m curious whether the 5090 package was not following UL requirements.

Would that make them even more liable?

Part of me believes that the blame here is probably on the manufacturers and that this isn’t a problem with Nvidia corporate.

replies(1): >>44469882 #
5. ryao ◴[] No.44469723[source]
Has anyone made 12VHPWR cables that replace the 12 little wires with 2 large gauge wires yet? That would prevent the wires from becoming unbalanced, which should preempt the melting connector problem.

As a bonus, if the gauge is large enough, the cable would actually cool the connectors, although that should not be necessary since the failure appears to be caused by overloaded wires dumping heat into the connector as they overheat.

replies(3): >>44469751 #>>44469763 #>>44470380 #
6. bobmcnamara ◴[] No.44469751[source]
Or 12 strands in a single sheath so it's not overly rigid.
7. alright2565 ◴[] No.44469763[source]
Might help a little bit, by heatsinking the contacts better, but the problem is the contact resistance, not the wire resistance. The connector itself dangerously heats up.

Or at least I think so? Was that a different 12VHPWR scandal?

replies(4): >>44469791 #>>44469795 #>>44469854 #>>44470734 #
8. lukeschlather ◴[] No.44469778[source]
Also, like, I kind of want to play with these things, but also I'm not sure I want a computer that uses 500W+ in my house, let alone just a GPU.

I might actually be happy to buy one of these things, at the inflated price, and run it at half voltage or something... but I can't tell if that is going to fix these concerns or they're just bad cards.

replies(2): >>44469918 #>>44470831 #
9. ◴[] No.44469791{3}[source]
10. bobmcnamara ◴[] No.44469795{3}[source]
Contact resistance is a problem.

Another problem is when the connector is angled, several of the pins may not make contact, shoving all the power through as few as one wire. A common bus would help this but the contact resistance in this case is still bad.

replies(1): >>44469801 #
11. ryao ◴[] No.44469801{4}[source]
A common bus that is not also overheating would cool the overheating contact(s).
replies(1): >>44469809 #
12. alright2565 ◴[] No.44469809{5}[source]
It would help, but my intuition is that the thin steel of the contact would not move the heat fast enough to make a significant difference. Only way to really know is to test it.
13. ryao ◴[] No.44469854{3}[source]
I thought that the contact resistance caused the unbalanced wires, which then overheat alongside the connector, giving the connector’s heat nowhere to go.
14. ◴[] No.44469882{3}[source]
15. wasabinator ◴[] No.44469918[source]
It's not the voltage, it's the current you'd want to halve. The wire gauge required to carry power is dependent on the current load. It's why when i first saw these new connectors and the loads they were being tasked with it was a wtf moment for me. Better to just avoid them in the first place though.
replies(1): >>44470000 #
16. autobodie ◴[] No.44469983[source]
GamersNexus ftw as always
17. dietr1ch ◴[] No.44470000{3}[source]
It's crazy, you don't even need to know about electricity after you see a thermal camera on them operating at full load. I'm surprised they can be sold to the general public, the reports of cables melting plus the high temps should be enough to force a recall.
18. AzN1337c0d3r ◴[] No.44470380[source]
They don't just specify 12 smaller cables for nothing if 2 larger ones will do. There are concerns here with mechanical compatibility (12 wires have smaller allowable bend radius than 2 larger ones with the same ampacity).
replies(1): >>44470568 #
19. kuschku ◴[] No.44470568{3}[source]
One option is to use two very wide, thin insulated copper sheets as cable. Still has a good bend radius in one dimension, but is able to sink a lot of power.
20. chris11 ◴[] No.44470734{3}[source]
I think it's both contact and wire resistance.

It is technically possible to solder a new connector on. LTT did that in a video. https://www.youtube.com/watch?v=WzwrLLg1RR4

replies(1): >>44474738 #
21. izacus ◴[] No.44470831[source]
With 5080 using 300W, talking about 500W is a bit of an exaggeration, isn't it?
replies(1): >>44474656 #
22. dreamcompiler ◴[] No.44472796[source]
To emphasize this point, go outside at noon in the summer and mark off a square meter on the sidewalk. That square of concrete is receiving about 1000w from the sun.

Now imagine a magnifying glass that big (or more practically a fresnel lens) concentrating all that light into one square inch. That's a lot of power. When copper connections don't work perfectly they have nonzero resistance, and the current running through them turns into heat by I^2R.

23. lukeschlather ◴[] No.44474656{3}[source]
I'm talking about the 5090 which is 575W.
replies(1): >>44475133 #
24. ryao ◴[] No.44474738{4}[source]
Uneven abnormal contact resistance is what causes the wires to become unbalanced, and then the remaining ones whose contacts have low resistance have huge currents pushed through them, causing them to overheat due to wire resistance. I am not sure if it is possible to have perfect contact resistance in all systems.
25. izacus ◴[] No.44475133{4}[source]
But why are you talking about it? It's a hugely niche hardware which is a tiny % of nVidia cards out there. It's deliberately outsized and you wouldn't put it in 99% of gaming PCs.

And yet you speak of it like it's a representative model. Do you also use a Hummer EV to measure all EVs?

replies(1): >>44475718 #
26. lukeschlather ◴[] No.44475718{5}[source]
I am interested in buying hardware that can run the full DeepSeek R1 locally. I don't think it's a particularly good idea, but I've contemplated an array of 5090s.

If I were interested in using an EV to haul particularly heavy loads, I might be interested in the Hummer EV and have similar questions that might sound ridiculous.