The backend GPU fabric is the inter-GPU fabric, and it supports two topologies: standalone switch, and multi-tier leaf and spine switches.
The standalone switch fabric is simpler and easier to implement. This fabric does not require complex traffic engineering features to guarantee performance, and it delivers the lowest switching latency performance. However, this fabric's scalability is limited to the number of switch ports.
The multi-tier fabric provides the performance and scalability for expansion; however, it requires specific GenAI networking features to be configured on the fabric and additional customization to guarantee optimum performance.
There are four distinct inter-GPU topologies: standalone, leaf and spine, pure rail, and rail optimized.
This guide discusses and describes how to modify three different AI blueprint examples:
- OOB management fabric
- Frontend fabric
- Backend (GPU) fabric
Each blueprint addresses specific fabrics and their functions