NSX-T is currently unicast only so no considerations around multicast support in the underlay, however, there are two modes (head end replication where the originating host sends a separate copy to each receiving host, and two-tier, where a single host is designated per subnet to handle local replication to other hosts on that subnet). The two tier mode is the default and generally recommended way to handle this, but there's a good description of both starting on page 18 of the VMware® NSX-T Reference Design
Thanks, so Head-End Replication mode is similar to Unicast Mode for NSX-V. If the Underlay Physical Network is L2 designed, i.e. all the VTEP Ips are on the same subnet, the BUM traffic is replicated to all ESX hosts that has member of N-VDS (Logical Switch VNI). For the Unknown Unicast Packets, the ESX host asks the Controller about this Mac address while at the same time forwarding this packet to other ESX hosts similar to broadcast or multicast packet (Page 23). This provides a "Fallback Mechanism" for Mac learning even for the rare case of Controllers not available, this mac learning behaviour could make the traffic flow although with some additional replication overhead.
The Two-Tier hiearchical replication (flooding mode) is the recommended way two optimize the CPU and reduce unnecessary replicationtraffic of BUM traffic. This explains why it the default mode. This looks like the Hybrid Mode, only the difference is local replication is again Unicast, while hybrid mode uses Multicast for local replication. If all the VTEPs have the Same IP Subnet, then Two-Tier and Head-End is same.
One possible best practice may be to design the Physical Network as L3 CLOS Fabric because of CPU and traffic optimizations, this again similar to best practice recommendation of NSX-V. This becomes more important as the scale of the network increases.