We run some benchmarks, then i have 'numerical' results.
I opened a ticket on the Freescale's support and they replied me that probably some kernel module was not included, then the higher speed.
I think that it is quite strange for benchmarks with float .. is it possible that a kernel module slows down by half the numerical benchmark??
Of course, this higher speed is welcome but if 'tomorrow' i will recompile the image and i will loose thise speed, i will not know why and how to restore the previous state.