Does anyone know how accurate the code execution profiling is running in SIL mode?

I have been trying to run some general code execution profiling on my PC (i7-8850 @ 2.6GHz) to try out some code generation optimizations. The measured execution times seem to be inconsistent at best when comparing multiple runs. Does anyone know of any documentation that specifies how accurate the execution times can be expected to be run to run or any best practices to get an accurate setup? I have seen the average time be as much as 20% different, the max time is pretty much completely worthless. The model setup is running the same inputs each time. I have tried short and long runs to try and smooth out the average. I am trying to gauge how useful this tool might be or if I should not waste my time on it and spend the extra time to run the code on the real HW (PIL is not an option). All thoughts and comments are welcome. Thanks!