NVIDIA’s connector story: EPS vs. 12VHPWR Connector – Unfortunately, good doesn’t always win, but evil wins more and more often (insights)

The Trick with the Smart Adapter

Since no native power supply cables were available, every board partner had to include a special adapter with the new Ada cards. Depending on the required power, these were initially adapters with three (450 watts) or four (600 watts) 8-pin plug headers for the 6+2 power supply cables. Theoretically, one could operate the card with just a single 6+2-pin cable, and this was feasible in practice. However, understanding how the adapter worked was key. NVIDIA hadn’t even disclosed the logic of the adapter to board partners and power supply manufacturers, so it came down to my lab, logical thinking, and some trial and error. At least now two sideband channels were connected.

Before the launch, I noticed that the power limit of the RTX 4090 FE automatically decreased when a cable was removed. This was frustrating because NVIDIA’s PCAT tool for measuring power consumption only works with three rails. Luckily, I had a well-stocked cable box and quickly found a Y-splitter. Or even two… It worked well, even with 2 cables and a hefty 600 watts.

NVIDIA probably counted on customer naivety (except for plugging in, but that’s another story). Therefore, it must be a simple logic circuit (AND gate) that connects the two pins of the 12VHPWR accordingly. So, I exposed all the lines and saw exactly what I expected from my initial measurements between the cables. Without any cables connected, the six 12V and six ground pins in the connector are already bridged, creating a single voltage rail. This eliminates the ability to detect missing rails, as was previously possible on graphics card boards when a connector was missing. And that’s why, with a bit of trickery, it’s unfortunately possible to operate the whole thing with just one cable. One of the two sense pins per 6+2 socket was not directly connected to any of the lines, nor to the first two sideband contacts in the 12VHPWR responsible for TBP detection. However, the resistance wasn’t infinite, but within the range typical for IC inputs. Aha, there we have it!

But let’s take another look at my previously exposed cable mess: What do we see in detail? For one, only one 12-volt and ground line per socket leads to the 12VHPWR, which isn’t a big deal due to the cable’s short length. Technically, the upper one of the two sense pins is directly connected to ground in the socket. Then it gets interesting! A total of 5 lines lead to the first 8-pin socket: the two from the sense pins of the 12VHPWR connector and three more cables, each leading to one of the other connectors.

This implies that a small logic circuit must be built into the first socket, elegantly and unnoticeably powered through the 12V line. Since we already know that all lines are bridged on a single rail at the 12VHPWR, Cable #1 doesn’t even need to be connected to the power supply, as long as at least one other is connected. Once you figure that out, you can draw a block diagram for clarity:

First Damages – Adapter with Soldering Errors

This clarifies how NVIDIA’s “miracle adapter” works. Well, if it works at all. Astron apparently still had to practice soldering. Right after the launch, the first damage reports hit social media. We also had a genuine defect with a scorched connection, and our investigation prompted NVIDIA to respond, though initially only within their board partner circles. NVIDIA’s quick and decisive response to the preliminary reporting was surprising, and should be acknowledged, even if damage and cost limitation played a significant role.

The problematic adapter came from Astron, the company that led its design. This leaves a bitter taste, as their engineers should have known better. Such soldering excesses from thick to thin, and crimped lines without real strain relief in the connector, are not only nonsensical and grossly negligent but technically prohibited as they contradict all electrical engineering rules. NVIDIA has not yet publicly commented, but they have contacted all affected board partners. Having the ability to directly ask responsible individuals in large companies, I might capture some statements.

 

I’m pretty sure NVIDIA itself didn’t know exactly what the supplier was doing inside the encapsulated connector. Otherwise, they would never have approved this design. Peeling back the layers gradually reveals something better left unseen:

Four thick 14AWG WIRES distribute over six contacts, with the outer lines soldered to one pin each, and the middle ones to two pins each. The solder base is a mere 0.2 mm thin copper foil, 2 mm wide per incoming line, resulting in 4 mm per pair for the middle connections. Soldering one or two 14AWG lines onto this is quite bold. Confucius might say, “What’s encapsulated isn’t visible, and what you can’t see, can’t break.”

But carefully removing the surrounding layer immediately tears the thin foil. This shows that bending the cables at the connector in the housing can cause such damage. Loose contacts or unsafe bridges, and increased resistance as a result of such actions, are extremely dangerous at these current strengths and quickly lead to the errors we’ve seen above. And now we see why the outer contacts of the connector are particularly affected. The scorched socket on the card is merely the consequential damage.

But one can further dissect the connector to see what’s inside. This detail is also highly interesting for root cause analysis. We see that the contacts in the connector are interconnected. If the outer two lines break, the entire current flows in the middle over the remaining two lines, then redistributing in the connector. However, this “foil” is thin and no replacement for real 14AWG. It’s obvious that it would get very hot…

Initially, I thought the cause would be easy to isolate. How wrong we all were, I’ve already explained in the article about the quality of headers and listed all the error sources encountered so far. But I’ll return to that later.