{"id":"leaf-switch","level":"datacenter","name":"Leaf switch and endpoint network","shortName":"Leaf switch","role":"Connects compute or service endpoints to the broader data-center fabric.","description":"Leaf switches are the access layer of a common data-center network. Servers or network adapters attach to leaf ports, and uplinks connect the leaf to spine switches. AI installations may separate compute traffic, storage traffic, and management traffic into different fabrics because their requirements differ. The atlas exposes these roles as selectable networks instead of coloring every cable as one universal data path. A rack's top-of-rack management switch is not automatically the same thing as its high-performance training fabric. Endpoint speed, uplink capacity, routing, and congestion control together determine effective communication. This level explains how multiple racks become a cluster while preserving the distinction between local scale-up links and cross-system networking.","specs":[{"label":"Topology role","value":"Endpoint access","context":"Connects hosts or adapters toward the spine layer."},{"label":"Traffic examples","value":"Compute, storage, management","context":"May use separate physical or logical fabrics."},{"label":"Capacity question","value":"Downlinks vs uplinks","context":"Oversubscription depends on the actual design."}],"interfaces":[{"name":"Compute and service adapters","direction":"in","signal":"Endpoint packets"},{"name":"Spine uplinks","direction":"out","signal":"Traffic to other leaves and systems"}],"tradeoff":"More endpoint capacity supports larger clusters. Insufficient uplink capacity can create contention.","whyAI":"Distributed training requires tensors to cross system boundaries. The access fabric is the first step of that communication path.","related":["spine-switch","optics","rack","storage","nccl","network-adapter"],"sources":[{"title":"NVIDIA — NVL72 AI Factory Network Logical Architecture","url":"https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory/latest/network-logical-architecture.html"}],"scale":"Rack-to-row network layer; topology is schematic","kind":"leaf-switch","mechanism":["Attach endpoints to leaf downlink ports.","Forward remote traffic toward an available spine path.","Deliver returning traffic to the appropriate local endpoint."],"misconception":"Compute, storage, and out-of-band management networks are distinct roles, even when all use Ethernet somewhere.","check":{"question":"What is a leaf switch’s main topology role?","options":["Connect endpoints to the larger fabric","Store every model checkpoint","Convert coolant into electricity"],"answer":0,"explanation":"A leaf provides the endpoint-facing network layer and uplinks toward other parts of the fabric."},"branch":"Data-center fabric","subbranch":"Endpoint access","science":[{"title":"Oversubscription ratio","equation":"O = Σ B_downlinks / Σ B_uplinks","explanation":"When aggregate endpoint capacity exceeds uplink capacity, simultaneous remote traffic can contend at the access layer. The relevant traffic pattern still determines actual congestion.","assumptions":"Directional port-capacity accounting with comparable payload rates; a ratio alone does not specify all fabric behavior."}],"history":[],"prerequisites":["network-adapter"],"learningObjective":"Calculate a directional oversubscription ratio and state what traffic pattern exposes it.","engineeringExample":"Synthetic leaf has eight 100 GB/s downlinks and four 100 GB/s uplinks. Downlink capacity totals 800 GB/s and uplink capacity 400 GB/s, giving 2:1 oversubscription. All attached endpoints cannot simultaneously send their full 800 GB/s through those uplinks.","evidenceNotes":"These are hypothetical comparable one-direction payload capacities. The ratio is a port-budget bound, not a prediction of local traffic, collective completion time, congestion control or application throughput.","labId":"library","scope":"Use the execution investigation for the shared tensor-to-memory-to-result thread; no current standalone calculator models this lesson’s full mechanism.","investigationId":"execution","labScope":"Use the execution investigation for the shared tensor-to-memory-to-result thread; no current standalone calculator models this lesson’s full mechanism."}