The simplest test I can think of (`docker run busybox true`) takes 0.38s wall clock, pretty far out from your statement even when allowing for it stopping as well.
I've not seen millisecond startup times since pre-1.0 on any of my machines (there was a post on the mailing list about the mailing list about this slowdown a while back). When/where have you seen this advertised?
Edit: unless we're excluding the time for the implicit `create` from this, which doesn't seem like a very helpful metric.