Yeah, certainly having "V"GPRs in a "scalar ISA" is a bit off. In graphics parlance, the term really just refers to the lack of things like dot products and matrix multiply instructions where the compiler has to line numbers into "per-thread" vectors for the instructions to work out. Turns out you can save a lot of work by not having that kind of ISA, and just moving to ones where you operate on number registers (which are then "vector"'d to be thread-count wide).

Wouldn't be graphics unless we overloaded a piece of terminology 20 times.