They are full 256 bit wide now, that's how it can claim a 2x float perf. improvement. It also doesn't have an "AVX" offset.