Edge sizing tends to get settled the way a restaurant gets picked when everyone's hungry: whoever names one first wins. Somebody says Large, nobody objects, and it goes into the design document with no reasoning attached.
Large is often fine. The trouble is that the size is only half of one decision, and placement is the half that bites later.
What the Size Is Actually Committing You To
The VCF design library recommends Large or X-Large Edge nodes and says Large is the best balance for most deployments, particularly when the Edge nodes are used only for north-south traffic. X-Large is aimed at larger-scale builds where many load balancer instances will run. If the Supervisor is going to use the NSX load balancer, Large is the floor. So the useful question isn't which size sounds nicest. It's how many load balancer instances this Edge cluster will ever carry.
Then comes the part people skip. An Edge node reserves 100% of its memory by default and carries High CPU shares. Every node you size up is memory the host can no longer hand to anything else. Multiply by the nodes per host and by the hosts you'd need to survive a failure, and the Edge cluster's real footprint shows up.
I'm deliberately not putting vCPU and memory figures for each form factor in this post. They change between releases, and they belong in your design document with the version number next to them.
Dedicated, Collapsed, and the Bare Metal Question
The design library recommends a dedicated vSphere cluster to host nothing but Edge nodes, and Edges can run as VMs or on bare metal. Running them in a collapsed cluster alongside workload VMs is possible, but it has costs. I'd start from the dedicated cluster and make the collapsed model argue its case.
The collapsed model fails in a specific, boring way. When other workload VMs in the same cluster also carry memory reservations, the Edge reservations can stop VMs from powering on, and vSphere HA may be unable to restart VMs after a host failure. It works fine until the day a host dies, which is the day you wanted it to work.
Bare metal is supported, with strings attached. The nodes need manual registration to NSX Manager, the NICs and CPUs have to be on the hardware compatibility list, and the docs themselves call it a more challenging deployment with limited topology use cases. That's a throughput conversation, and I'd rather not wing the numbers here.
An Edge node is a VM that behaves like a router. Routers don't enjoy being treated like VMs.
Counting Nodes to NICs
Edge nodes per host should be limited to the physical NICs available. Each node wants a dedicated pair of pNICs, and fp-eth0 and fp-eth1 get mapped to complement each other, so each one is active on one physical NIC and standby on the other. Four nodes on a host with two uplinks isn't a design, it's a hope.
The network span matters too. Keep Edge nodes local to the rack, with BGP peering to the top-of-rack switches in that same rack. Stretching an Edge's reach across racks to save a few ports tends to cost much more later in troubleshooting time.
MTU is the other quiet one. The Host and Edge TEP minimum is 1700 (Geneve is supported as low as 1600, and the extra leaves some headroom). The Tier-0 external interface should match the physical peering interface, and the global gateway MTU sits 200 below that: 9000 outside means 8800 inside.
The Tier-0 Rule That Changed Under a Lot of People
In the centralized connectivity model, two Edge nodes is the minimum, and the recommended Tier-0 mode is Active/Active, which scales out to eight nodes. The segment model makes the same recommendation, runs stateful services on Active/Standby Tier-1 gateways, and leaves the other Tier-1s in DR-only mode.
Where it gets tangled is NAT. VCF Automation All Apps organizations and the Supervisor depend on Default Outbound NAT (Auto-SNAT), and that needs stateful services, which means the Centralized Transit Gateway has to be Active/Standby. In earlier releases the transit gateway shared the Tier-0's HA mode. In 9.1, a newly created transit gateway has its own, and the shared mode is deprecated, so an Active/Active Tier-0 and an Active/Standby transit gateway can coexist. Gateways carried over from an upgrade may still be in shared mode, so check them. And if your Edge design came from a rule of thumb learned on an earlier build, check that too.
One more Supervisor note while we're here: starting with 9.1, the vSphere Client no longer deploys a Supervisor on classic NSX segment networking. That path is API only.
What Gets Fixed by the Edge Design
- Routing model. EBGP between the Tier-0 and the physical network, with each Edge node peering with two separate physical devices. BFD on, which lets you leave the BGP timers at their 60 and 180 second defaults.
- Room to grow. An uplink peering subnet of /28 or larger is what lets an Active/Active Tier-0 scale to eight Edge nodes later. For an Active/Standby Tier-0 a smaller subnet is fine, since it's capped at two Edge nodes. The /28 is what buys the scale-out.
- The load balancer choice. Avi is one-to-one with an NSX Manager and has to be integrated before Supervisor activation, or the Supervisor quietly falls back to the NSX load balancer.
Maintenance Is Where Placement Gets Tested
This is the one I'd hand to anyone running an upgrade window. When a Best Effort Restart compute policy is in place, vMotion can't be used to evacuate an Edge node during host maintenance. You build the host groups yourself before running the Network Connectivity wizard. With host group affinity selected, the wizard or the Edge cluster workflow then creates the VM groups, the must rule, the vCenter tag, and the BER policy itself. The docs are explicit that you shouldn't hand-build those four.
Even then, putting a host into maintenance mode yourself shuts the Edge down whether a healthy peer is available or not. Someone has to confirm the other node is carrying traffic first. The runbook should say who.
One more habit worth dropping. The CPU graphs for an Edge node in the vCenter UI aren't reliable. Use the NSX Edge CLI, API, or UI, or VCF Operations.
Questions I Ask Before Anyone Picks an Edge Size
How many load balancer instances will run on these Edge nodes, now or in the next release cycle?
Does anything in the design need Default Outbound NAT, and what HA mode is the transit gateway in?
How many physical NICs does each Edge host really have, and how many nodes does that allow?
If one host in the Edge cluster is lost, can the rest still power on every reserved node?
Who confirms the peer Edge is carrying traffic before a host goes into maintenance?
If your Edge cluster was sized by whoever spoke first, I'm happy to look at it. DM me and we'll find a time that works for both of us, remote.