More anecdata: On my workloads, which involves moving large discrete image frames as part of a video inference ML pipeline through a message broker and other processing pipelines, de-mitigating decreased system loads ~25% on average. Some workloads up to 30% depending on inference model. So yes noticeable.