Requirements for using Tensorcores

[Bananenbrotesser]

Hello,

i would like to use the https://github.com/stanfordnqp/jaxwell repo for my FDFD simulations. I want them to be as fast as possible.
The main computational part is the iterative solver for a large system of linear equations. (https://github.com/stanfordnqp/jaxwell/blob/main/jaxwell/cocg.py)

All elements are named tuples (https://github.com/stanfordnqp/jaxwell/blob/main/jaxwell/vecfield.py).

My idea is to modify the code to use the tensorcores of the A100 GPU. Currently tensorboard shows the created kernels as "not eligible for tensorcores"

What requirements do they have? Are there better ideas to achieve a speedup?