# Leaf switch and endpoint network

Canonical: https://semiaifoundry.com/learn/atlas/leaf-switch/
Revision: 2026-10-07
Rights: https://semiaifoundry.com/ai-atlas/licenses.html

Data-center fabric · Atlas v1.3.0

# Leaf switch and endpoint network

Connects compute or service endpoints to the broader data-center fabric.

[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#leaf-switch)Rack-to-row network layer; topology is schematic

Learning objectiveCalculate a directional oversubscription ratio and state what traffic pattern exposes it.

## The mechanism

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.

1.  Attach endpoints to leaf downlink ports.
2.  Forward remote traffic toward an available spine path.
3.  Deliver returning traffic to the appropriate local endpoint.

## Science and assumptions

### Oversubscription ratio

$$
O=\frac{\sum_i B_{\mathrm{downlink},i}}{\sum_j B_{\mathrm{uplink},j}}
$$

Oversubscription is summed downlink capacity divided by summed uplink capacity, using the same direction and bandwidth convention.

Symbols and units

O

Directional oversubscription ratio

dimensionless

B\_downlink,i

Comparable payload capacity of downlink i

bytes/s

B\_uplink,j

Comparable payload capacity of uplink j

bytes/s

i,j

Included directional port indices

integer indices

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.

## Engineering example

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.

## Distinctions and tradeoffs

Compute, storage, and out-of-band management networks are distinct roles, even when all use Ethernet somewhere.

More endpoint capacity supports larger clusters. Insufficient uplink capacity can create contention.

Distributed training requires tensors to cross system boundaries. The access fabric is the first step of that communication path.

## Specifications and interfaces

Topology role

Endpoint access

Connects hosts or adapters toward the spine layer.

Traffic examples

Compute, storage, management

May use separate physical or logical fabrics.

Capacity question

Downlinks vs uplinks

Oversubscription depends on the actual design.

### Interfaces

Compute and service adapters Input

Endpoint packets

Spine uplinks Output

Traffic to other leaves and systems

## Check your understanding

### What is a leaf switch’s main topology role?

1.  Connect endpoints to the larger fabric
2.  Store every model checkpoint
3.  Convert coolant into electricity

Answer and reasoning

Connect endpoints to the larger fabric

A leaf provides the endpoint-facing network layer and uplinks toward other parts of the fabric.

## Evidence and scope

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.

Use the execution investigation for the shared tensor-to-memory-to-result thread; no current standalone calculator models this lesson’s full mechanism.

## Sources

-   [NVIDIA — NVL72 AI Factory Network Logical Architecture](https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory/latest/network-logical-architecture.html)

[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/leaf-switch-e818b606ed058bca.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)

## Continue exploring

### Foundations

-   [Network interface adapter](https://semiaifoundry.com/learn/atlas/network-adapter/)

### Related lessons

-   [Spine switching and fabric topology](https://semiaifoundry.com/learn/atlas/spine-switch/)
-   [Optical transceiver and fiber link](https://semiaifoundry.com/learn/atlas/optics/)
-   [Integrated accelerator rack: GB200 NVL72 example](https://semiaifoundry.com/learn/atlas/rack/)
-   [Dataset and checkpoint storage hierarchy](https://semiaifoundry.com/learn/atlas/storage/)
-   [Collective communication library](https://semiaifoundry.com/learn/atlas/nccl/)