I'd argue that calling the new matrix multiplication unit they added to the GPU cores a neural engine instead of a tensor processing unit is a branding error that will lead to confusion.
The existing neural engine's function is to maximize power efficiency, not flexible performance on models of any size.
I'd argue that Apple's definition of "neural engine" was entirely different from what the greater desktop, edge and datacenter markets already considered a "neural engine" to be.
The existing neural engine's function is to maximize power efficiency, not flexible performance on models of any size.