## Limitations of Traditional Leaf-Spine Architecture

Traditional data centers use a leaf-spine two-tier network architecture, but face severe challenges in intelligent computing scenarios:

- **High bandwidth convergence ratio**: Traditional designs typically use 1:3 to 1:5 convergence ratios, which is seriously insufficient in AI training scenarios
- **Low cross-host communication efficiency**: Servers on different leaf switches must forward traffic through the spine layer
- **Complex RDMA deployment**: Requires dedicated network configuration to ensure lossless RDMA transmission

## Non-Blocking Architecture

### What is Non-Blocking?

Non-blocking architecture means any port to any other port in the network can communicate simultaneously without congestion due to insufficient internal switch bandwidth.

### Key Features

- **1:1 convergence ratio**: Each downlink port bandwidth equals uplink port bandwidth
- **Any-point-to-point connections**: Any server can communicate at full speed simultaneously
- **Supports large-scale RDMA**: Supports AI training clusters with thousands of GPUs

## Technology Implementation Paths

### Path 1: Upgrade to Higher Bandwidth

Upgrade leaf-spine architecture to 400G/800G port density, reducing convergence ratio by increasing bandwidth.

### Path 2: Clos Architecture Extension

Build larger-scale CLOS networks by increasing the number of leaf and spine switches.

### Path 3: Dedicated AI Network

Use dedicated networks for AI training scenarios, such as InfiniBand or RoCEv2 networks, separate from general business networks.