Hi ! I saw on the NVidia site this interesting new library for the Matrix Multiplication called CUTLASS.
They have done a great job from a performance point of view there, it will be nice for several applications especially Machine Learning.
The library is versatile, it is built on purpose to implement in details specializations for your application.
Tiling and Software Pipelining are used to increase the performance while taking in account the architecture and the memory hierarchy:
Check out this other post too about Tensor Cores: https://devblogs.nvidia.com/parallelforall/programming-tensor-cores-cuda-9/
Probably in a few weeks or months it will be integrated in a few major libraries.