Timo Birnschein's Temp Logger Looks to Save His NVIDIA GeForce RTX 5090 From a 12VHPWR Flame-Out
Overheating problems on the high-amperage 12V input of NVIDIA's latest RTX graphics cards has buyers, like Birnschien, worried.
Engineer Timo Birnschein is looking to protect his investment in a high-end NVIDIA RTX 5000-series graphics card — by monitoring its 12VHPWR input for signs it's overheating.
"Instead of whining about the power of my GPU and NVIDIA's design issues," Birnschein explains, "I decided [I'd] rather implement a thermal watchdog on the connectors. This is a very quick one: [an] Arduino with a few 4.7k resistors and 100k thermistors. On the Windows side, a Python script running with administrator rights checks the temperature of the three thermocouples and logs them. Once at least one passes 100 degrees Celsius, the PC automatically shuts down — regardless of data loss. Way better than destroying $2000 in hardware or more."
The problem Birnschein is trying to solve: the extremely high power draw of NVIDIA's top-end graphics cards, with the flagship GeForce RTX 5090 specified at a "thermal design profile" of 575W — just for the GPU, not counting the rest of the system into which it's installed. Those kind of power envelopes, which come with the suggestion to have a system power supply of at least 950W, are far more than the PCI Express slot can handle, and even more than standard PCIe power cables. NVIDIA's solution: 12VHPWR, a high-amperage 12V power connector on the edge of the card.
Sadly, top-end GPUs from both the new GeForce RTX 5000-series and older 4000-series have been known to scorch, melt, and even flame out — owing, analysis has suggested, to a lack of balancing on the power input's multiple pins. If the cable isn't quite clicked home, or if some of the pins are less than perfect in either surface finish or oxidation, the card will draw additional current through the better pins — resulting in overheating to the point of melting and scorching.
The problem is so pronounced on the company's latest generation that third-party manufacturers have stepped in with cooled cable connectors — adding either heatsinks or active fans to try to tame the heat. Birnschein's approach is more reactive than proactive: monitoring the connector's temperature and pulling the plug if things start to get toasty.
"The project is ongoing," Birnschein says, pointing to some bugs which still need to be ironed out including stalls in the Python script graphing temperatures. "I used an Arduino Nano [compatible development board]. A0-A7 [are] read by the firmware. A 4.7k resistor goes from each pin to +5V. The thermistor is connected directly to the pin and GND as a voltage divider. That's it. I use the Steinert algorithm to calculate the temperature and dump it onto the serial port. Currently, only A0 to A3 are read and communicated."
Interested parties can follow the project's progress on Birnschein's Hackaday.io page.