Network Topology in AI Cluster
Frontend
Backend
Management

AI Cluster Design Challenges
Non-blocking and Lossless – Congestion management (DCQCN – ECN + PFC)
Oversubscription and bandwidth considerations
Latency
Load balancing
Physical considerations (cabling, cooling, power, etc.)

Traffic Patterns in AI Cluster
North-south and east-west traffic
Comparison to traditional datacenter traffic patterns

Consequences and Implications of a Bad Design

Transport Protocols in AI Cluster
TCP for frontend
RoCEv2 for backend

Traffic Flow in an AI Cluster

Rail Optimized Inter-GPU Communication
When leaf only, when spine, when super spine, etc.

AI Cluster Configuration (NX-OS for routed fabric)

AI Cluster Operational Models
On-prem – Nexus Dashboard: how to create an AI fabric 
Cloud-managed – Hyperfabric AI Fabric demo (web demo)