Agreed with "usually". When you model pipelines, you must know the cache behaviour of your workload, otherwise it is a waste of time. But when you do, in order to approach theoretical machine limits, you do need airtight core resource utilization in your kinner loops.