3 ms·
The API may not be copyrightable, but the widely-used official libraries (cuDNN, cuBLAS, OptiX, etc) are, so in practice I think they can at least say that anyt
by jsheard 3y ago
The API may not be copyrightable, but the widely-used official libraries (cuDNN, cuBLAS, OptiX, etc) are, so in practice I think they can at least say that anything which uses those libraries can't be run on competitors hardware.
- zozbot234 3y agoAIUI, these libraries are generally implemented as shared objects that the resulting binary has to link to. So it should be quite possible to reimplement the interface that they expose, moreover something like ROCM has to do that anyway in order to compile HIP source code from scratch. Looks like that's what https://github.com/ROCm/hipDNN/ https://github.com/ROCm/hipDNN/ does.
- jsheard 3y agoPossible, yes, but reimplementing the CUDA runtime and the official libraries and getting them up to par with the originals is a much bigger task than just doing the former and running Nvidias libraries in it. AMD did try to do a complete reimplementation of cuDNN with hipDNN but "last commit 5 years ago" doesn't inspire much confidence in it being competitive with cuDNN proper.
- imtringued 3y agoThey don't have any tensor cores so there is not much reason for them to work on this.