Typically we recommend to run minimum 3 nodes (grid initial-size 3 --> 3 initial nodes) in the grid for production use. This is because grid etcd quorum so you can safely lose one node and other two still continue to run without issues. The initial nodes for the data replication cluster on etcd and any further nodes act as pure proxies. So the grid operates perfectly as long as there is a majority of the initial nodes available.
For RAM requirements on nodes it mostly up to the application requirements. Platform components (agent, etcd, ipam, cadvisor) on the nodes consume typically only around 200 megs.
Master itself doesn't consume that much resources in normal cases. Master is actually mostly sitting idle and does it's hard work only during deployments. So typical 1G RAM, 2cpu machine is enough. Unless you are planning to run high number of services and instances (containers) and especially if the services are producing lot's of log data. If that is the case then you'd need bigger machine for master to make it cope with high load of log data streaming in from the agents.