I built this PC about four years ago with an Intel Core i9-12900K, MSI PRO Z690-A WiFi motherboard, 32GB Corsair DDR5-6200 memory, Corsair RM1000X 1000W power supply, and a Corsair 5000X case. It originally used an RTX 3090, which gradually developed artifacts and eventually stopped displaying video. A replacement RTX 4090 lasted longer but eventually failed in the same way. Its RMA replacement is a refurbished Gigabyte Windforce OC RTX 5090, and the packaging had clearly been opened and damaged.
The system has also occasionally shown a message saying that the previous overclock settings failed before entering the UEFI menu. The memory is rated for 6200MHz with XMP enabled, but the system sometimes falls back to 4800MHz while leaving the XMP profile selected. I normally exit without changing anything and it boots successfully.
After installing the RTX 5090, the motherboard's PCIe retention latch did not seem to fully lock. The card appears to be pushed against the case's expansion-slot bracket before the latch reaches its normal locked position. The GPU is currently powered through two PCIe cables using their daisy-chained connectors and the included 12VHPWR adapter, although I am replacing that setup with a proper PCIe 5.1 12V-2x6 cable.
The first boot produced a DRAM debug light and no video, which was fixed by reseating the memory. The system now reaches the operating system, but after a few minutes the display cuts out, the VGA debug light comes on, and the fans suddenly run at extremely high speed. Initially the fans stopped after a few seconds, but now they continue until I force the computer off. What should I check first, and could the motherboard, memory settings, power supply, or PCIe connection be responsible for these repeated failures?
5 Answers
Use a dedicated, PSU-compatible 12V-2x6 cable designed for that exact Corsair power supply rather than relying on daisy-chained connectors and an adapter. Make sure every connector is fully inserted and that the cable is not sharply bent near the GPU. The 1000W capacity should generally be adequate, but after two graphics cards have failed, I would also test with another known-good PSU and inspect the motherboard, PCIe slot, and power connectors for scorching, looseness, or physical damage.
The combination of video loss, a VGA debug light, and fans running at full speed usually means the system has lost communication with the GPU or the GPU has crashed hard. Test the card outside the case if possible, with the motherboard supported properly and the GPU fully inserted. If BIOS is current, the slot is secure, memory passes testing, and a proper power cable does not help, the refurbished card or the motherboard becomes much more suspicious.
Update the motherboard BIOS before drawing conclusions about the new card. A Z690 board predates this generation of GPU, and newer BIOS versions can improve PCIe compatibility and DDR5 training. After updating, load BIOS defaults and test the system with XMP disabled first. If it becomes stable, manually try a lower memory speed such as 6000MHz rather than immediately using 6200MHz XMP.
The graphics card needs to be fully seated and the retention clip should actually lock. If the case bracket is stopping the card from going in far enough, loosen the motherboard or bracket screws temporarily and test the fit. Also check for anything around the slot, such as a chipset heatsink or case alignment issue. Do not keep running the system with a partially connected PCIe slot.
Test the memory separately instead of assuming the graphics card is the only problem. Run a proper bootable memory test, and if needed use a more demanding Windows-based test such as TestMem5 or Karhu. Test with one stick at a time and with XMP off. Repeated DRAM training failures and VGA failures can be symptoms of unstable memory or firmware rather than two unrelated dead graphics cards.
You can check the PCIe link setting in the BIOS under the motherboard's PCIe or chipset configuration. If Auto or Gen 5 is unstable, forcing the graphics slot to Gen 4 is a useful troubleshooting step and usually has negligible real-world performance impact.