bench: avoid peakflops register dependencies

Created by: melven

Improves results on Intel Westmere which was missing a factor of 2 before due to latency/register dependencies in the peak flops benchmark.