>Here's another theory - maybe the permute instruction does more than just swizzle/replicate?
>If the instruction were extended to support simple modifiers (absolute value, negate,
>maybe even scaling by powers of 2), then that could concievably be counted as another 4 FLOPS.

This new Altivec unit is lifted from APU and is different from the traditional Motorola version. If PPE's Altivec unit could do 12 FLOPS, then so does APU's vector unit. But APUs are rated at 8 FLOPS/s.