Could be dynamic frequency scaling. To minimize the impact of it when benchmarking one can pin the process to a single core and warm it up before running the benchmark.