Problem with RTX card on Ubuntu 18.04 server headless

I run Tensorflow computations with a Nvidia RTX 2080. The driver setup was made rigorously following the Tensorflow web page documentation. So I have installed the drivers 418.67, Cuda 10.0 and Tensorflow 1.13.1.
I need to run a long computation with Tensorflow, and I have noticed that this computation is on average much slower than on my laptop equipped with a Nvidia Geforce 1070. Typically, the RTX 2080 computes the 3-4 first loops much faster (30-40% faster), then slows down twice slower than the GTX 1070, then after a while, regains its speed, slows down, and so on.
I have tried to deactivate, reactivate persistenced in nvidia-smi, but it does not change anything. The card does not get hot (the fans are not even activated).
I tried to run a benchmark test (https://github.com/keras-team/keras/blob/master/examples/mnist_mlp.py) and with that test the RTX 2080 card keeps a consistant much faster speed than the GTX 1070. So I do not know what is going on.
It seems that on my particular computation the RTX card goes to sleep, then wakes up, then so on.
I need to run my simulation for several hours, so I would greatly appreciate some help.

Please run nvidia-bug-report.sh as root when the issue appears and attach the resulting .gz file to your post. Hovering the mouse over an existing post of yours will reveal a paperclip icon.
https://devtalk.nvidia.com/default/topic/1043347/announcements/attaching-files-to-forum-topics-posts/