• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

Possible workaround for NVidia connectors - use poor quality cables

qxp

Joined
Oct 27, 2024
Messages
151 (1.39/day)
It might be that the workaround to NVidia connector problems is exactly the opposite of what people are recommending, and short, high quality cables are bad for 4090/5090. As discussed in other threads, the problem with 4090/5090 happens because the wires are supposed to carry large current in parallel, but the contact resistance can vary significantly, resulting in uneven current distribution in the wires.

This effect is greater for short, high quality cables, which is what the MODDIY cable was: good thick wires with gold plated contacts.

To see why, consider a short cable, 1 foot in length (30cm), using 16 gauge wires. The individual wire resistance is 0.004 Ohms. The contact resistance can vary between 0.006 Ohms to say 0.016 Ohms.
Then, the maximum variation between good and bad is (2*0.06+0.004) to (2*0.006+0.004) which is a factor of 7.75. A current difference of a factor of 7.75 results in a power dissipation difference of a factor 60 - meltdown.

Now, suppose you are using a similar cable, but 3 feet in length (1m), with the same 16 gauge wires. Then, the individual wire resistance is 0.018 Ohms, and the ratio between good and bad contacts is at most (2*0.06+0.012)/(2*0.006+0.012)=5.55

If the same 3 feet cable was made using cheaper 18 gauge wires, then ratio between good and bad contacts is at most (2*0.06+0.019)/(2*0.006+0.019)=4.48. For 20 gauge wires, we get (2*0.06+0.030)/(2*0.006+0.030)=3.57.

To summarize, the wire resistance acts to buffer the difference in contact resistance and balance the current. Longer cables with thinner wires will have less variation in resistance for each individual power lane compared to short cables with thicker wires.

We can't go too low in wire gauge because 20 gauge is already questionable, even for chassis wiring, if you include the possible factor of 3.57 current spike:


Using wires with higher-temperature insulation (like PTFE) will allow for the use of thinner wire with higher resistance.

There should be similar effect for contact plating: gold-plated contact should be able to achieve lower resistance per contact and higher current spike compared to nickel or tin-plated ones, but I don't know by how much.

Of course, a better long-term solution is an improved connector:

 
Joined
Oct 22, 2020
Messages
57 (0.04/day)
Also I suppose that a cable delibirately-crafted with high-resistance wire but still quality gold-plated contact would give even more predictable balance. Fixed predictable resistance from wire and minimize resistance addition abd heating-near the connector.
 
  • Like
Reactions: qxp

qxp

Joined
Oct 27, 2024
Messages
151 (1.39/day)
Also I suppose that a cable delibirately-crafted with high-resistance wire but still quality gold-plated contact would give even more predictable balance. Fixed predictable resistance from wire and minimize resistance addition abd heating-near the connector.
Indeed.

Usually the power supply rails are specified to be no tighter than 5%, so GPU 12V could be 11.4V instead for a voltage drop of 600mV. We get about 8.33A per wire for 600W load at 12V, which means that the wire+contacts resistance should be no more than 0.072 Ohms. Reserving 0.01 Ohms per contact, this leaves at most 0.05 Ohms for the wire, which would result in nominal power dissipation of around 3.5W per wire - not bad.

The simplest solution would be to splice a piece of nichrome wire of the same diameter is copper. 16 gauge nichrome wire is 0.25 Ohm/ft, so about 60mm or 2.4" would provide 0.04 Ohms of resistance, assuming each crimp uses up about 6mm or 1/4".


Would be a good idea to cover it with a high-temperature heat-shrink - polyimide or PTFE.

Ideally the splices would be staggered to spread the 3.5W heating load.
 
Top