I found a http://newsroom.intel.com/servlet/JiveServlet/download/1202-20-2558/Infos_zu_Intel_Demos.pdf where the performance of Knights Ferry/Corner? is compared to 2 x 2.93 Xeon X5670. It looks like it's 1.25x faster in HPCC GUPS (highly parallel, but low SIMD affinity), 2x in energy-seismic, 2.25x in academic matrix math benchmark, 3x in CERN Track Fitter benchmark, 3.25x in CFD - Lattice Boltzmann Model, and 5.25x in SP Monte Carlo (highly parallel, high SIMD affinity, use of transcendental functions). Looks like Intel Braunschweig helped. The chart on the right also looks reminiscent of old Larrabee graphs.
