It’s a long story about restoring a physically broken NVIDIA RTX 5090. Probably, some other video card should have been thrown away, but this one is too expensive.
Bottom line damaged RTX 5090 ended up in a video by Northwest Repairs, a company that specializes in difficult repair cases. The video card was cracked and showed «signs of life». The repair attempt turned into a series of adventures with the discovery of new faults and their successful overcoming. Cracks in the PCB can interrupt signal paths hidden inside the textolite and lead to seemingly random and unexplained failures.
The first surprise was revealed after removing the cooling system. The memory contact plate didn’t touch the evaporation chamber properly, so there was virtually no heat dissipation. The power modules and other components seemed to be intact, so the next step was to test the power supply.
Even in standby mode, the video card was consuming about 5A — quite a lot, but maybe normal for a 5090. The current was increased to 8 A to get the video card to start up: the lights came on, the fans spun, but there was no image, despite the monitor’s LED signaling power on. This meant that the video card was working, but not producing an image, which led to the assumption of problems with the video memory.
A computer with a processor that had a built-in video card helped identify a flaw in one particular memory chip. The video card was unable to complete the DRAM initialization sequence — a sign of chip damage, likely a damaged solder joint. The 2GB Samsung GDDR7 module was unsoldered and re-soldered to the board.
However, the card’s fans immediately accelerated to maximum speed, indicating that the repair is still far from complete. Also, the PCIe interface suddenly failed, which could have been due to shutdown attempts and hidden damage that led to further problems. Graphics processor was subjected to rebolting (re-soldering) — careful work with many contacts that requires precision and skill.
The changes led to a new problem — a lack of memory, probably due to heat. Thermal imaging and sequential testing revealed a faulty module — the one that had been re-soldered. After replacing it completely, the short circuit was eliminated, but the system still did not recognize the chip because the PEX to PCIe protocol was still not working.
At this point, the initial board defect probably got worse due to all the heat. Only a few phases of the VRM power supply were actually working. The 12V and Vcore lines were working, but the digital “Driver ON” signals were not reaching half of the VRM controllers. The trail led into the thickness of the board. Since the signal was not distributed between the phases, a simple jumper was made to restore the broken path. This restored full power, but the PEX did not work. Another jumper was added to cover the missing PCIe power-up signal — this worked and the card demonstrated a video signal.
But the hopes of completing the repair were still not realized. PCIe detection was not working again, the fans were at maximum, and one phase of the memory was not turning on. After checking, it turned out that the PEX coming from the 3.3V bus of the PCIe slot was not reaching its destination, which actually caused this phase to shut down. The final fix was to run a third path to supply 3.3V directly to this memory power-on signal. After that, everything worked.
The PCIe connection finally became stable, the power was balanced across all phases, and the video card passed a full round of stress testing in benchmarks and in games. This was one of the cases when keeping her expensive video card was not only possible, but absolutely justified.
Spelling error report
The following text will be sent to our editors: