Modular/Mojo is faster than NVIDIA's libraries on their own chips, and open source instead of binary blob. See the 4 part series that culimates in https://www.modular.com/blog/matrix-multiplication-on-blackw... for Blackwell for example.

thanks