They do benefit from batching; up to a 50% performance improvement, in my experience.

That might seem small compared to LLMs, but it isn't small in absolute terms.