Quote Originally Posted by JF-AMD View Post
2. That work is done in 128-bit FP for the most part
4. A Bulldozer 12-core would have 12 128-bit FP engines
Rather no, Sandra uses the AES instruction set. AES instructions are internally decoded into Doubles. i.e. 2 µops. These two ops are routed to the first FMAC port, where there is not only a FMAC Unit but also a tiny, little AES Unit connected and to the second FMAC port, where there is also some permutation unit (named FPXBR or XBAR). Thus, the FMACs are never in use.
So you cannot count 2 Units per module. Yes, there are two Units, but both are used simultaneously for only 1 (double decoded) instruction.

Attachment 119478