GPU: NVIDIA GeForce RTX 5090 32GB
Motherboard: Gigabyte Z890 GAMING X WIFI7 - CPU: Intel Core Ultra (Arrow Lake) -
RAM: 64GB Crucial PRO DDR5-4800 (JEDEC defaults)
PSU: Super Flower Leadex VII Platinum PRO 1200W
OS: Ubuntu 24.04
(Noble), kernel 6.17.0-29-generic (HWE)
Driver: nvidia-driver-580-open 580.159.03 - Workload: Boltzgen protein structure prediction and sustained CUDA compute for hours/days
I’ve been fighting GSP-related hard crashes on my RTX 5090 since mid-May. Identical to other reports in this thread — abrupt hard reboot or complete freeze (no SSH, no display, mouse LED
off). Post-crash boot shows MCE hardware errors and BERT error records. No Xid errors in journal because the crash is too sudden for the kernel to log them. Journal simply stops dead mid-write.
After extensive A/B testing, I’ve identified two key findings that may help others: (1) pcie_aspm=off is a significant partial mitigation, and (2) motherboard BIOS version has a
dramatic impact on crash frequency — upgrading my Gigabyte Z890 BIOS from F18 to F21a made things catastrophically worse. (3) Currently running F20a and so far so good!
Thanks to this thread. I was able to get the pcie_aspm=off to stabilize things! . But I am still keeping my fingers crossed that the new Bios is able to improve the stability. Also a slight aside the nvidia kernel driver 595.71.05 had a segfault in libcuda.so.595.71.05 during repeated PyTorch CUDA forward passes (confirmed upstream:
Segmentation fault on RTX 5090 with CUDA 13 during repeated PyTorch CUDA forward passes ). Rolled back to 580.159.03.
My second post here , but excited to have this platform to benefit from and share my experiences