It's one thing to have an API and hardware support for such things, it's a different thing entirely to have support from the major ML stacks (pytorch/tensorflow etc).