Quote Originally Posted by informal View Post
Many people came to a conclusion,primarily from C11.5 numbers,that Zambezi will have very strong FP units.A total score of almost 11 pts tells us that per one single FMAC(in shared MT mode) we have roughly 30% more computational power than one whole Thuban core,roughly at the same clock.When Zambezi doesn't see load across all cores,FMACs can be "paired up" (per module) so that in single or poorly threaded fp workloads we have a lot higher fp resources,on top of those 30%. Now on top of all this comes new Turbo boost in poorly threaded workloads.All things summed up,at least in legacy (or FMA tuned) fp code,Zambezi will be a big step up from today's cores,yes even versus SB.
The two (2) 128-bit FMACs per Module will only be paired to one 256-bit FMAC when we will have AVX.

If we only have one FP instruction per Module, then one 128-bit FMAC can have all the resources (FP Sheduller etc).