"Harmful" doesn't mean "so dead in the water that it doesn't actually have performance gains". Intel wouldn't have integrated SIMD if it didn't work at all; that's not the criticism. Rather, the claim is that the performance gains are not well justified by the technical debt they bring to the architecture. The performance gains are not as good as they could be with the vector approach, which has better dynamic and static code density, doesn't horribly proliferate the instruction set, and allows for tuning without recompilation of code to use different instructions.