Concepts

concept

The mental models the rest of the knowledge base builds on. Read these first; every layer, workload and path page links back here.

• The nine-layer model

  ---

  The stack from AI application to GPU silicon, and how Kubernetes participates in each layer without implementing most of them.

• Hardware/software co-design

  ---

  Why the slowest rank sets the pace, a worked 256-GPU example, and how to correlate signals across layers.

• Parallelism

  ---

  DP, FSDP, TP, PP, CP and EP: what each one splits, the traffic it creates, and where it must be placed.

• Life of a training step

  ---

  One gradient followed through all nine layers, and what breaks at each hop.

• Workload taxonomy

  ---

  Training, serving, pipelines and interactive work, and which infrastructure each one stresses.

• Related models

  ---

  Other public models of the AI stack, and how this one compares.

• Glossary

  ---

  Terms and acronyms used across the site.

Suggested reading order

1. The nine-layer model
2. Parallelism, then each strategy page
3. Workload taxonomy
4. Hardware/software co-design