What I found interesting in THG preview is Sisoft Sandra FMA results. Haswell does perform better than IB/SB but not nearly good enough when you look at the "theoretical" boost FMA could bring- they gained only around 18% of FP performance Vs 256bit AVX codepath. They mention that the FMA results are latency bound so this could be potential bottleneck . This could mean that Haswell-E/EP won't be making huge performance jumps(~2x) in spec fp tests when compared to IB Xeon derivatives. Maybe this is why AMD has not opted to expand their FlexFP unit to native 2x256bit FMA pipes, they expected that intel won't be getting such a massive lead with Haswell's implementation of FMA.