Bitfury chips actually have
756 cores.
Sounds like they're using "unrolled" (i.e. crappy) cores.
actually the doc says double hash or 768
Not exactly, they say they say each core has two units which "share work" across one "job". Whatever that means. Presumably a nonce range, so one chip would work on nonce 0xFFFF0001 while the other would work on 0x0001FFFF, or some other distribution.
Regardless of how they share work each "sub core" completes one hash per clock cycle so it isn't like they are only partial work engines.
So 4 * 96 * 2 = 768 hashes per clock cycle.
At 550 Mhz that is 768 * 0.550 = 442.4 GH/s.