Hopper had 60 TF FP64, Blackwell has 45 TF, and Rubin has 33 TF.
It is pretty clear that Nvidia is sunsetting FP64 support, and they are selling a story that no serious computational scientist I know believes, namely that you can use low precision operations to emulate higher precision.
That's a surprisingly large overhead. I've not measured that large an impact on AMD, particularly for compute heavy.
Did you profile at all? And have you observed if it's not compute-bound? If it's memory or IO bound it can be due to other virtualization overheads, such as memory encryption.
It doesn't go to zero, however!