let's say we have powerful machines and our use case is less write while more search.
can I deploy multi instaces only cover search responsibility while one instance cover write&search responsibility on one machine?
Yes, you could do this (with shadow replicas), but you should be asking yourself if you want to do that. You can leave a single node per machine and it will happily read and write your indexes in an optimal way, including handling the caching of results for the node. By introducing a second node on the same physical hardware, you may actually decrease your overall performance.