The backend (GPU) fabric provides the actual communication fabric. There are two types of topologies that are provided as part of the SFM GPU blueprints.
- Standalone
- Rail optimized
This guide describes the modification of a 64 GPU blueprint using a single Z9664F-ON Dell PowerSwitch.
- From the main SFM dashboard, click Blueprints and then click Export.
- Edit the exported blueprint. Depending on the type of GPU blueprint chosen, you need to modify one or two Excel worksheet.
The example being used for this AI guide requires editing only one Excel workbook.
For this 64 GPU Layer 2 fabric, you can modify the following Excel worksheets:
- Overview - This worksheet deals with the overall description of the blueprint.
- Nodes - This workbook contains the specific node information.
- Breakouts - This workbook contains information about any breakout interface that is configured.
- Fabrics - This workbook lists the fabric parameters, such as ASN and others.
- Networks - This workbook lists the specific networks configured on each interface.
- Downlinks - This workbook lists the networks that are networks that are configured on each interface.
- InventoryDevices - This workbook lists the number of end devices with their respective GPUs. The AI blueprints use the Dell PowerEdge XE9680 with 8 GPUs as the device.
- InventoryDevicesNICs - This workbook lists the number of connections from the switch to the Dell PowerEdge XE9680 NICs. The AI blueprints use 400 GbE speed as the link connection.
Note: The Z9664F-ON is a 400GbE native switch. The Z9864F-ON is an 800GbE native switch. Whenever the Z9864F-ON switch is used as part of any of the AI blueprints, the InventoryDevicesNICs connections are split into two physical connections. For example, there will be 64 physical devices to NIC when using the Z9664F-ON, and 128 physical devices to NIC when using the Z9864F-ON switch.
The process of editing the Excel workbook is outside the scope of this guide. For this information, see the Dell SmartFabric Manager for SONiC user guide .
After editing the workbook, save the modified Excel workbook using a unique name, making sure that the Overview column of the new blueprint is also renamed with a different name. Otherwise, the Import process will fail.
- Select the new excel spreadsheet and Import the file.
- To deploy the newly updated GPU blueprint, click CREATE INFRASTRUCTURE.
When you create the infrastructure, you can define specific blueprint parameters. Items such as switch IP address management, LAGs, specific networks, and other blueprint characteristics are entered during this phase of the blueprint deployment.
- To deploy the new updated GPU blueprint, click DEPLOY.