NVIDIA 12VHPWR Testing & Discussion

So I thought I would make a seperate thread here to not derail the 5000 series thread any further. After a lot of feedback on my initial video, I've decided to take this further and keep investigating…

You can see the initial video here:



Next up to test, everyone is saying I gave the connector too much thermal mass and “cheated” in a way so Im doing another quick video using the same 16AWG wires on both sides. Personally I think the connector will still do better than the PCB but we will see.

I originally chose the 12AWG wires to stabilise the pins between each other as soldering a connector by hand isn’t easy and going from 6x16AWG to 3x12AWG seemed fair…



I have other things to explore also but ill keep those cards close to my chest for now…

Thanks for all the constructive feedback on here also, Youtube hasn’t been anywhere near as nice

EDIT 1: Follow Up Video 1



EDIT 2: Follow Up Video 2

 

Some images from testing tonight, 3 x 30 minute cycles holding sustained 1300W around 165-171c. The hottest wire temperature yet

Also an overview of heat distribution towards the bus bars as this seems to be a contentious point for people (I'm open to ideas how you connect this much stuff without using bus bars...

Will have a short video coming on this soon, another video will follow later testing another aspect of connector durability

Things to be changed:

Increase wire length between connector and bus bars (more representative of a typical run to a power supply and creates more distance from the bus bars)
Switch to lead free solder (higher melting point)
Asking around for someone to design a proper multi layer PCB

 

Wow, getting braver

171C and nothing melted?

That seems like it would be well over what any of those jacketed wires would be rated for, like 2x?

Were / how were you able to check for damage and compare before and after?

Are 1300W loads are still useful? I think people are more interested in knowing what failure modes exist at 600W - though transients may factor in, 600W that spikes to 750W for instance might be interesting

 

Nothing melted but starting to see signs of thermal stress on the housing (this is still the original female connector which has been abused a lot now).

The wires are labeled as rated for 80c but obviously that’s not their melting point…

1300W isn’t useful obviously but Im just exploring the outer limits of the connectors capabilities and what can be learnt from the setup overall in regard to improving it.

I can definitely look at programming the loads with ramping rates / bursts / endurance cycles also. Next up is insertion durability tests (giggity) I think which I have some plans around how to do it but will swap out to all new connectors and start over fresh to be fair on it.

 

Ready for the next batch!

 

wtf?

So where is the failure on these cards? Not the connector?

 

Exploring many possibilities still obviously, but my thoughts so far on things to explore…

• Poor pin contacts causing abnormality high resistance/very localised heating at the contact points which heats faster than it can be dissipated through wires etc
• Garbage China cables using 18AWG wires / poor QA with crimping
• Counterfeit / knockoff connector housings and connectors with poor tolerances, materials etc
• PCB trace designs that aren’t generous enough and contribute to excessive temperature rise / heating

I think Derbauer already showed an example of this with 20a on a single pin which was a good example of a failed cable, it was interesting that the heating was at the power supply end though…

 

You might want to get some safety perspex

 

Or safety three pluggers.

 

Poor pin contact and thus a hot joint I think would be the biggest issue as well as your other points being factors.

a couple of other videos had described the variations and versions of pin contact/connectors used, all critical failures in design are due to a perfect storm scenario.

It must be said well done on your testing, with a more electrical engineering/electronics methodology to your testing it really cuts out the guessing game most content creators are doing.

Perhaps you could find a dead (just for parts) RTX 4xxx series card on eBay for cheap and use the PCB and connector as a test point.

 

Derbauer's video showed 20a on one wire because Nvidia stopped putting proper load sensing into their pcb's.

 

The Asus Astral cards have per pin power sense but it looks like they are the only ones. They however don’t current limit, it’s still connected to a bus after and then two load resistors. So it’s only a warning system.

It is starting to look like a PSU problem with thin cables and bad crimps/third party connectors.

 

At least there's a warning system, allows you to shut down and try to get a good even connection across the connector, or at least a more even one. No idea why nvidia decided they would take what would be a couple of $'s worth of parts and pcb design out of the protections knowing full well they are pulling 600w out of a 600w rated cable/connector.

 

Bus bars will naturally be a heat sink. no getting around that. You could use terminal strips instead.

But I think that is not the point anyway. you're testing the connectors.

the temp of the wires is but a data point, easier to measure than individual current per wire. I'd argue current per wire/pin is more important (will show imbalance which the 5000 cards seem to have troubles with), but it's not as flashy as a nice thermal pic. temp will of course dictate when the connectors begin to melt, insulation melting/burning. but it's a side effect of too much current on that pin/wire.

from what I've seen, poor quality connectors (small wire gauge, bad pin contact, pins being out of place in the connector etc.) will cause a current imbalance across the 6-pins/wires. The fact the 4xxx and 5xxx cards have no balancing circuitry then causes it to become a runaway issue.

what'd be cool imho to test, would be to remove the plastic shells, and test the pin to pin connections individually (no need to run all 6, run them individually at reduced power loads - 100W instead of the 600W for all six), have them properly inserted, partially inserted, cocked to the side etc. it's after all not the plastic shell that gets hot. I've not seen anyone do this level of testing (outside of Molex - the manufacturer)

 

I sort of get arguments both for and against it. On one hand it is nice to have it there but on the other as a designer, you have been tasked to cut BOMs and shrink the PCB… you have a connector dumping into a single 12v bus for the entire card so why would you add extra complexity / take up real estate to mitigate a problem with someone elses PSU / cable etc. Not saying one is right or acceptable but it is also easy to armchair this stuff without seeing both sides.

One issue with PCs is nothing is really harmonised and its everyone for themselves. If you were designing a system holistically from end to end then of course you would account for such things.

 

Thanks for the suggestion to depin them as this is something I could do potentially. Current I have been checking periodically using the current clamp but I haven’t been remotely happy with its accuracy to be honest. My meter can tolerate up to 20A for brief periods but disconnecting things also causes issues by disrupting the setup of course

The thermal views are just being used to look for variations in wire temperature which seems to give signs of uneven current distribution on the pins. I do need to settle on a power level in next tests though and close the temperature scaling in so I can see finer variations.

 

depending how much you want to invest, you can get high current current shunts pretty cheaply, then you measure them with a volt meter, and convert the milli-volts read to Amperes - as the voltage drop over the shunt is directly related to the current flow (as the shunt is a known resistance).

heck even some car audio style current meters would do the job cheap enough 6 wouldn't be too difficult to afford and you could show 6 real time displays..

(apologies if I'm telling you to suck eggs, not sure on your level of electrical expertise)

 

Currently discussing with some techs about doing a multilayer PCB design for the connector and could possibly implement the shunts as a part of that which more closely replicates a graphics card PCB also on individual pins then bring them together into a common trace. Would prefer it on the other side though if possible. I expected the current clamp accuracy to be better so ill definitely explore this

EDIT: Found some 30A shunts that claim 0.5% accuracy for a decent price… worth a shot before i start the next round of testing

 

try a terrible old psu with terrible terminals. (with safety equipment)
ie components that will fail possibly more easily.
===
or don't
==
point being there are always a certain number of failures of assembly line products.
whether it is the psu/bad terminals or something else. a defective card (from any manufacturer or brand)
or even a defective plug that has not met manufacturing tolerances off the assembly line.
'known' examples of the cable issue are very very rare <====
==
i hope you have fun testing various things though. just your original vid was reassuring in terms of the capacity (over-capacity!) of the connector/wires.

 

Oh I am having fun haha, I originally overbuilt everything so hopefully the only failure will ever be the connector itself. Was shocked I ran out of load though…. *smirk