Optical network topologies
Compare dual-fabric, spine–leaf and industrial branching without confusing hops with latency.
Concept documentation. Not a production specification. API examples are illustrative, not hosted services.
A topology is a set of relationships
A network diagram records devices and their connections. The NEXORA lab represents those relationships as an undirected graph: nodes are devices and links connect pairs of nodes. This intentionally small model supports deterministic inspection and path tracing. It does not simulate Ethernet forwarding, optical circuit setup, transport congestion control or production routing protocols.
AI Training Cluster
Six compute nodes attach to two optical fabric nodes. A selected compute pair can communicate through an intermediate fabric node, and another structural route may also exist. The model illustrates parallel connectivity rather than certifying a particular resilience scheme. Real operation would need a definition of how traffic uses those links and what happens after a failure.
Hyperscale Data Center
Two spine nodes connect to four leaf nodes. Each illustrated server attaches to a leaf. Communication between servers on different leaves can traverse leaf, spine and leaf relationships. This highlights a hierarchy distinct from the smaller cluster view. The drawing does not assert nonblocking capacity, equal-cost multipath operation or any oversubscription ratio.
Industrial Edge Network
A controller connects to two gateways, each serving a small branch of devices. Robot, sensing and vision labels describe endpoint roles. They do not imply industrial protocol support, deterministic control or functional safety. The topology is useful for explaining why a gateway is a boundary between compute and distributed equipment.
Read the highlighted path
Select distinct source and destination nodes, then highlight a route. The browser performs a shortest-hop graph search. Every consecutive node pair must correspond to an actual link in the selected fixture. A packet then traverses that same route. Changing presets clears the previous path because node identifiers and connection relationships differ.
If two nodes have no connecting route in a graph, path search should return an empty result rather than inventing a line. The shipped presets are connected, while the network unit tests separately exercise disconnected and invalid endpoints. A production tool would also need to distinguish administratively disabled links from unknown or stale observations.
What the diagram cannot tell you
Hop count is not latency, and a drawn line is not an optical budget. Distance, supported optics, queuing, endpoint behavior and operating mode all influence real outcomes. Path diversity is also different from independent physical routing. Read data center interconnects for channel boundaries and AI cluster connectivity for workload considerations. Use the lab to learn and inspect, not to dimension or certify a real network.