A40 Link Capability register was changed and nvidia-smi can't see the card

I am doing a power cycle test a server with 2 A40 cards. The GPU sometime will downgraded at PCIE Gen 3…

And nvidia-smi only can see one GPU card:

But lspci can show both:

The support file attached below:
nvidia-bug-report.log.gz (545.7 KB)

any idea to solve this issue is welcome