The benchmarking tool being used in this post accounts for that using multiple runs for each invocation, together with a warmup run that is not included in the result metric.