That's highly dependent on access patterns and domain space.

When I was optimizing for render sorting on mobile hardware we knew we would never see 2k+ drawcalls and cache aware tuning gave us huge end-user facing benefits.