If we're talking low level SIMD, like opcode level, I'm really struggling to see the use case for interpreted bytecode. The cost of type checking operands to dynamically dispatch down a SIMD path would almost certainly outweigh the savings of the SIMD path itself.

JIT is different because in function-level JIT, you can check types just once at the opening of the function, then you stay on the SIMD happy path for the rest of the function. And in AOT, you may able to elide the checks entirely.

There is certainly a space for higher level SIMD functionality. Numpy is one example.