That's covered at the bottom of the post (tensor-valued vs scalar-valued). It would make some of the math kernels faster, since numpy has a bunch of specializat

That's covered at the bottom of the post (tensor-valued vs scalar-valued). It would make some of the math kernels faster, since numpy has a bunch of specializations built-in. But it would not speed up any of the connective "glue" bits that are still going to be slow. And it would still probably involve a bunch of allocations.