Each predicate will belong to exactly 1 group. Data replication is only between nodes in the same group. There is no cross-group replication.
E.g. you could have 4 nodes, split across 2 groups (2 nodes in each group). Replication would occur within each group, so that each edge in the graph is duplicated between 2 nodes.
The passage on the website is really talking about situations where you need HA (high availability). By setting up servers in groups of 3 (i.e. 3 per group), one server doing down would leave 2 remaining servers. 2 remaining servers could likely still handle normal load.
If HA is critical for you, then you should run at least 3 dgraphzeros, and 3 dgraphs. If you are able to have more nodes, then you could run 3 dgraphzeros and 6 dgraphs split between 2 groups.
If all nodes in one group fail, then you will experience downtime until the nodes are brought back online.